AgentLog

xAI has rebuilt the Grok Bot product page around running multiple Bots as "AI teammates". The updated page pitches messaging Bots like colleagues on desktop or iOS, running several at once - one on a project, one on outbound, one on systems - working in parallel and collaborating where it makes sense, around the clock. "Show a Bot how it's done" saves a demonstrated workflow as a routine the Bot runs on its own next time, and "Connect the Bots" puts several Bots in the same thread so they pass work between themselves while you watch, instead of approving every step.

Most of this is not new. Multi-Bot parallel work, Bot-to-Bot coordination, routines, and getting sharper over time were all in the August 11 launch post, which the refreshed page still links as its launch announcement. Shared Team Bots followed in late September, and the Premium, Super and Ultra pricing shown now arrived with the X Subscription on October 12. What the refresh actually adds: a claim that Bots now "learn from each other", not just from their own user; a preset roster of Bot jobs to start from, including Sales Outbound, Talent Scout, Expense Manager and Chief of Staff; FAQ wording that puts the desktop app on macOS alongside Windows for the first time; and Grok Bot presented as generally available across Premium, Super, Ultra, Teams and Enterprise plans.

Cursor's October 16 changelog gives cloud agents a new output type: canvases. A cloud agent can now build an interactive canvas as part of its work, and you can share the finished canvas with your team by sending its link. The update lands two weeks after the Origin App Marketplace connected Cursor to more than 20 deployment, CI, code review, and monitoring apps.

The same changelog opens the Environments API in beta, which lets you programmatically create and manage the cloud environments your agents run in, and raises the ceiling on those environments: up to 300 repositories each, with large environments loading five times faster.

BAND, the agentic collaboration platform for multi-agent systems from Thenovi AI Ltd., launched BAND for Developers on October 15: a desktop application, alongside its terminal and CLI, that lets software engineers manage and collaborate with multiple coding agents at once. Instead of passing context between isolated agent sessions and coordinating handoffs by hand, developers get a visual view across their agent work - projects, agents, ownership, activity, blockers, questions, and progress - and can drop into a specific project or agent to answer a question, redirect it, or review its work, then step back to the broader picture.

The app is designed for workflows where specialized agents share one outcome: one coding agent implementing a feature, another testing it remotely, a third handling another part of the work. Agents can exchange context, divide responsibilities, and hand tasks to one another across supported harnesses, machines, and environments, without forcing teams to standardize on a single agent stack. Agents running locally can be connected with remote or centrally managed organizational capabilities. CEO and co-founder Arick Goomanovsky frames the shift as developers going "from working with one coding agent to leading teams of them, but the tools around those agents haven't evolved to support that."

Microsoft announced a set of Copilot Studio releases on October 15 that turn agents into persistent participants in shared business processes. An agent can now get its own organizational account with approved credentials and scoped permissions, so people can invoke it in Teams and Outlook, @mention it in Word, Excel, and PowerPoint comments, and hand off work by email. The agent keeps going between interactions, and its actions are attributed to the agent's own account rather than borrowed from a user, which gives administrators lifecycle, access, and audit controls.

Business process memory lets an agent carry user preferences and process context into each new cycle, so a quirk one person raises becomes shared context for the whole team; makers decide through instructions which records and policies are authoritative. Triggers in the GitHub Copilot harness start an agent on an external signal, like an arriving email, or on a set cadence, with no prompt required. Still to come: refinement recommendations that turn stable repeated steps into workflows and skills, and agents running on their own browser-enabled cloud computer.

The features are rolling out through Microsoft's Frontier early-access program. Microsoft names Ooredoo Qatar as an early customer, running an executive business-case review agent across Outlook, Teams, and Word.

Cognition launched Devin Oncall on October 15, a generally available agent that investigates alerts and bug reports and helps teams diagnose and resolve production incidents. Rather than stopping at a probable cause, the agent keeps working until the underlying issue is resolved - rolling back or scaling up when that is the right mitigation, or opening and testing a pull request before paging an engineer to review it. Per-organization skills are generated automatically: Devin Oncall regularly indexes the customer's environment and builds procedures for common debugging tasks, like calculating the customer blast radius of an issue or querying platform dependencies through integrations such as the Datadog MCP.

Cognition says Devin Oncall saturated the SREGym-Lite benchmark with 100% end-to-end success at the fastest time to mitigation, and the launch post leans on production evidence: at Reducto, Devin spotted concerning trends and opened an incident on its own before the company's alerts fired, and now handles most engineering alerts first; at RV Tech, the Rivian-Volkswagen joint venture, it extracts identifiers and timestamps from customer reports, pulls logs, and triages issues for review; at AngelList, it identified root causes 3.5x more often than the in-house harness it replaced and handled half of incidents end-to-end. Cognition says the agent already handles thousands of real production alerts every day.

MacPaw, the maker of CleanMyMac and Setapp, launched Eney on October 15, a proactive AI assistant for Mac and iPhone that works across your apps, files, and system. Instead of waiting for a prompt, Eney watches the Mac's context - the apps you use, files on disk, meetings on your calendar - and puts a prepared next step in front of you for approval: a call about to start, a meeting that needs a summary, a disk running out of space. The Eney Wand command (shake the cursor over text, Option+W, or double-tap Option) lets you point at anything on screen and hand it to the assistant without copy-pasting into a chat window.

Eney runs on a mix of local and cloud models, with custom models co-developed with Liquid AI for Apple hardware, and uses Mnemos, MacPaw's on-device memory system built on a knowledge graph; Mnemos is still limited to managing Eney's own memories. It can import ChatGPT archives, fix spelling and grammar in Mail, Slack, Notes, or the browser via Eney Spell, transcribe calls without a bot joining, and prepare notes, action items, and follow-up invitations. Mac maintenance features cover low disk space, memory pressure, crashing apps, and Downloads cleanup, and ready-made workflows include Morning Brief, Plan My Day, and Weekly Review. Eney works with Apple Mail, Calendar, and Reminders out of the box and connects to Gmail, Google Calendar, Google Drive, Slack, Notion, Jira, and Linear. It launches in eight languages including Ukrainian, and requires an M-series Mac on macOS 15 or later.

Pricing is credit-based: a 14-day free trial with 10,000 AI credits, then Starter at $14.99/month (10,000 credits), Boost at $29.99 (20,000), or Power at $49.99 (30,000). Setapp sign-in is supported but a Setapp membership is not required.

Hermes published stable 0.22.0 on October 14; its release notes are dated October 13. A fresh desktop install opens a guided setup chat on a free trial model, available once per operating-system user, so getting started does not require an API key. Hosted connectors bring apps such as Gmail, Linear and Notion into the agent, with several named accounts supported for each app. Local Models setup installs and manages llama.cpp on macOS, Windows and Linux, helping users choose an open model that fits their hardware.

The release also documents tested stable update channels, signed desktop packages, browser sign-in through 1Password, Bitwarden or Hermes's encrypted vault, and a Bot Screen that users can take over for login, two-factor authentication or CAPTCHA steps. This is the feature roundup for the whole development window, including 0.21.1 through 0.21.7, rather than a claim that every feature arrived after the last patch.

Bees publicly introduced its free, open-source desktop app on October 13, with downloads for macOS, Windows and Linux. Describe an outcome in plain language and it turns the work into stages on a shared board. Agents work through those stages, and a separate reviewer checks each result before it moves on. Bees supports Claude Code or Codex subscriptions, hosted providers and local models, plus recurring processes and local search over folders and mapped Google Drive locations. A private workspace starts without a Bees account.

The public introduction is not a new build: the latest stable 0.2.2 release was published October 7. Approval is off by default and must be enabled per stage to gate sends, posts, deletes or payments through tools. Mac packages are signed and notarized; Windows and Linux packages are unsigned. The 0.2 series does not migrate existing local data from 0.1.1, so those users need a backup before trying it.

Google's Gemini Apps help page now describes starting a task from the same place you start a chat: below the text box, click Task. The rename replaces the Spark name in the main interface while keeping existing tasks, threads and schedules, which still appear in the sidebar and on the Scheduled page. 9to5Google first spotted the change on October 12 as a server-side rollout on Android, where Spark history and schedules were combined into one page.

The requirements are unchanged: you need a personal Google Account (not work or school) with a Google AI Pro or Ultra subscription, Keep Activity on and Personal Intelligence opted in. Tasks remain unavailable in the European Economic Area, Nigeria, Switzerland and the United Kingdom, and Google still caps you at 15 tasks running at a time. The Android version of the same help page still uses the Spark wording, so the rollout has not finished everywhere.

On October 11, Yun-Ta Tsai, identified by TeslaNorth as a Tesla AI engineer, posted that "Grok @bot is enabled on all Cybercabs and the rest of the Model Y fleet." This brings access to the task-running Grok Bot into the vehicle interface, rather than just the conversational Grok assistant. The post does not name a software version, explain how to enable it, or extend the claim to other Tesla models.

AgentOne 1.1, announced on October 10, adds Scheduled Agents to its desktop app. You can set up recurring work in chat or enter a prompt and cron schedule manually. When a run is due, macOS, Windows or Linux starts the app through its own scheduler, and AgentOne begins a new chat using your existing default model, tools, extensions and settings. Prompts and schedules stay on your computer; there is no cloud scheduler.

OpenClaw published the stable 2026.10.1 release on October 10. MCP Apps can call their tools while their views remain open, with inline message confirmations and synchronized forms, files and context. Code Mode's store/load values now remain available across cells, replies and Gateway restarts. The release also fixes problems with updates, session continuation, model responses and message delivery.

OpenAI announced composer predictions in Codex on October 9. After Codex finishes a response, it can suggest a complete next message based on the current thread. Pressing Tab adds the suggestion to the composer, where you can review or edit it before sending. The beta is for adults aged 18 and older on personal ChatGPT Pro plans in supported regions, using local or SSH threads in the Codex desktop app with GPT-6 Astra or GPT-6.1 Sol. Predictions are on by default for eligible users and can be turned off under Settings > General > Composer > Show predictions.

Grok Bot's October 9 announcement says a Bot can now have its own email address to sign up for services, contact businesses, or schedule time with someone. A follow-up in the same thread says the feature is rolling out to users and that they can claim an inbox by asking their Bot or tagging @bot on X. Team admins must enable it for their team.

ChatGPT's October 5-9 Dots update says users can now create a dot in the mobile app, give it a name, customize its appearance, and connect plugins from their phone. An October 9 announcement on the OpenAI community forum specifies iOS and Android. Dots can also start work in Codex and follow up on existing threads, using context from ChatGPT conversations, Codex threads, and automations. They can review and edit ChatGPT Work automations, too.

At Gemini at Work on October 8, 2026, Google announced a Gemini agent for business work. Google says it can plan multi-step assignments, use skills and tools, connect to business systems, and keep context across sessions. Teams can create coworker agents with their own Workspace account, email address, calendar, Drive, and company-directory presence. Those agents act under their own identity and see only what the team shares with them. VentureBeat reports that the new agent is in private preview for selected customers, with no confirmed date for wider availability.

TechCrunch reported on October 8, 2026 that Cal AI co-founder Zach Yadegari's new startup, Persona, raised $10 million in a round led by Vine Ventures. Persona already has a free iMessage assistant beta, with a few thousand users according to Yadegari. Its website describes an agent that makes calls, books appointments, cancels subscriptions, and returns confirmations. The company also takes preorders for an optional $179 wearable band, which TechCrunch says is expected in December. October 8 is the date of the funding report, not the assistant's original launch: Yadegari introduced Persona in September.

Grok Bot's October 7, 2026 changelog says Bots can build slide decks with you, show sample slides to choose from, and deliver PowerPoint files or Google Slides. Bots can also send email with headings, bold text, and lists from a draft card, search Gmail's Spam folder and move a message back to the inbox, and notify a Slack user group from a Team Bot. The notes say @here and @channel still do not notify anyone.

Meta's Muse app now supports iPad as well as iPhone. The App Store release notes say "Muse is now available on iPad," and TechCrunch reported the expansion on October 7, 2026, about a month after Muse's mobile debut. The listing requires iPadOS 18.0 or later.

On October 6, 2026, Brett Adcock's Hark released Hark Pro widely. It is a personal agent for the web, iOS, and Android. It uses its own cloud computer, called Handoff, to log in to sites and finish tasks. The app is free, and Hark says every feature stays on the free plan. A $20 monthly plan allows twice the usage, and a $100 monthly plan allows ten times the usage.

nanoMuse is an early open-source personal agent with Android, desktop, and web clients, plus an iOS TestFlight beta. Its repository includes the apps and the relay that connects them under GPL-3.0-or-later. The project says it can use a phone's screen and a computer to carry out tasks, keep working after an app closes, and ask before actions that cannot be undone. The repository's October 6, 2026 release notes describe fixes to phone chat ownership, screen capture on Mac, Windows startup, and the relay's operator pages. Users can run their own relay instead of using the community service.

On October 5, 2026, Cohere introduced North 2, the next version of its enterprise agent platform. Agents can keep context across sessions instead of starting from scratch, and a new agent harness runs multi-step work. Admins can set spending caps, rate limits, and per-user quotas.

On October 5, 2026, Instinct began letting early access users add it to group chats, where a separate group Instinct works for everyone in the thread, even friends who have not joined Instinct. Founder Noah Shinn said your personal Instinct asks before it connects to a group's agent, the group agent cannot access your personal accounts, and your personal agent always asks before sharing information or taking any action. It reaches all users "soon," and you can ask your Instinct to put you on the list.

On October 5, 2026, 404 Media reported that Meta found security flaws in Muse in the weeks before launch, including at least one that could have let a normal user break an agent out of its isolated virtual machine and reach Meta's internal databases and services. The issues were raised to Mark Zuckerberg, and an internal post said a hardening push started on August 27, 2026, with teams working nights and weekends so the launch would not slip.

Perplexity's October 5, 2026 changelog adds Automations to Computer on the web. A task can run on a schedule or when something happens in a connected app, use Memory, Skills, and connectors, build on earlier runs, and pause for review before a consequential action. The same notes add inline charts, including TradingView charts for candlesticks, volume, and moving averages, and GPT-6.1 Sol for eligible paid subscribers in Perplexity and Computer, where it also powers Light effort.

On October 3, 2026, WIRED reported that Muse's own instructions tell it to keep a page for every person in the user's life, refreshed hourly, covering family, partners, friends, colleagues, collaborators, and people you follow. A researcher pulled those instructions through the chat interface, and Meta said the agent needs relationship context to be useful, gathering it from public information and what you choose to share.

On October 2, 2026, Meta published Muse Gadgets, open-source ESP32 firmware and a Linux SDK so people can connect Muse to their own hardware. It also built Muse Home Link, a USB-C stick based on the ESP32-C5 chip that puts Muse on a home Wi-Fi network. Meta made 5,000 of them and is giving them away free to subscribers in the United States with an active Muse plan, one each, with shipping in October.

OpenClaw released v2026.8.35 on October 2, 2026, a gateway-only update on its extended-stable branch, the project's long-term-support-style track. The highlight is GPT-6.1 Sol, plus safer updates and fixes so delegated replies stay complete. The release notes say this line is the August 2026 release plus later fixes, and that mainline was already on 2026.9.7 when this cut went out.

On October 1, 2026, Cursor said GLM 5.3 and GLM 5.3 Flash are available in the editor. Cursor's model docs describe GLM 5.3 as Z.ai's open-weight coding model after GLM 5.2, with a context window of 1 million tokens and Low, High, and Max reasoning settings. Cursor says GLM 5.3 Max is the best-scoring open-weight model on CursorBench 4.0, and GLM 5.3 is billed at the same per-token rates as GLM 5.2.

On October 1, 2026, xAI's Grok Bot account on X said bots can suggest ways to help without you asking. A follow-up in the same thread said your primary Bot will spot work it can take off your plate and offer to handle it. The change was rolling out over the next few hours, and those suggestions do not count against your usage.

On October 1, 2026, Perplexity announced 10 ready-made skills for Perplexity Computer, made for US American Express business card members who also have a Perplexity Enterprise subscription. The skills cover jobs such as cash-flow forecasts, tax-season prep, marketing drafts, and hiring. Members open Personal CFO from the connectors page, link a card through Plaid, and running a skill uses Computer credits.

On September 30, 2026, xAI's Grok Bot account on X said bots can hand coding tasks to Cursor, manage pull requests with GitHub and Origin plugins, and share video demos of what they build.

On September 30, 2026, 9to5Mac reported a Sensor Tower estimate that Muse had passed 5 million downloads, about 22 days after launch. Sensor Tower said ChatGPT, Grok, and Claude took 56, 103, and 492 days to reach that mark. Meta increased ad spending for Muse on September 16, and downloads rose 73% that day.

On September 30, 2026, Tailscale explained how Meta's Muse joins a Tailscale network as its own device, separate from the phone or laptop you already use. Once connected, Muse can see other machines on that network, work with self-hosted services, and handle a Tailscale SSH session. It only makes outbound connections, asks the first time it reaches a device, and you can allow that access once or on an ongoing basis and revoke it later.

On September 30, 2026, Perplexity CEO Aravind Srinivas said anyone can forward or CC [email protected] to hand the agent a task, with no Perplexity account. The agent finishes the work in the background and keeps the context from the email. Perplexity said those emailed tasks are free for a limited time.

Meta announced on September 29, 2026 that Muse is expanding beyond personal use to small businesses. The new version connects to the tools many small business owners already use, including Shopify, Stripe, QuickBooks, Notion, Slack, Zoom, Asana, Box, Canva, Figma, and more, plus a business's Instagram, Facebook Page, and ad account data. Meta says the goal is an agent that understands a business's whole operation, not just a stand-alone chatbot, so it can help owners run day to day tasks and find new customers. Muse for Small Business is free with usage limits, with a paid subscription for businesses that need more.

OpenAI announced Dots on September 29, 2026, at its DevDay conference. Dots are always-on personal agents built into ChatGPT and powered by GPT-6 Astra. Each Dot works from its own cloud computer that you can open to check its progress, keeps working in the background toward goals you give it, learns your preferences from your feedback, and connects to more than 4,000 apps through plugins. You can reach it from ChatGPT on desktop, web, and mobile, and from Slack and Teams. One Dot is included for Pro and Business Premium subscribers in eligible markets, rolling out in stages. Pro users in the EEA, Switzerland, and the UK are left out at launch, and Enterprise, Edu, and Healthcare workspaces get a beta once an admin turns it on. OpenAI says Dots can still make mistakes, and usage will not count against plan limits for the first month.

OpenClaw announced OpenClaw Enterprise on September 29, 2026, a free MIT-licensed control plane for running persistent agents on your own servers, with separate team spaces, sandboxes, and an audit log. The work started at OpenAI, then moved to the OpenClaw Foundation, and Red Hat and NVIDIA are helping build it. It is meant for internal pilots, with a 1.0 release planned later this year.

On September 29, 2026, at its HOOD Summit in Houston, Robinhood announced Robinhood Agents. It is an agent inside the Robinhood app that can research markets, build a strategy, and place trades for you around the clock. You give it a dedicated account. Reviewing each trade before it is placed starts turned on, and you can turn that off. Robinhood's own HOOD Summit disclosures say you assume all risk for trades the agent makes, and that Robinhood does not control, supervise, or audit the agent. It is coming soon for eligible U.S. customers.

Anthropic released Claude Sonnet 5.5 on September 28, 2026, the second model in its new Claude 5.5 family after Opus 5.5. It produces output more than 30% faster and costs up to 30% less per task, with pricing unchanged at $2 per million input tokens and $10 per million output tokens. Anthropic says it is especially strong at well scoped everyday tasks, bug fixes, and polished documents, slides, and spreadsheets, the kind of work Cowork hands off to Claude. Claude Desktop added support for it the day before, in version 2.9939.4.

xAI opened a public beta of Team Bots on September 28, 2026, a version of Grok Bot built for a whole team to share instead of one person. A Team Bot is set up around a shared role, like sales, product management, marketing, or data analytics, using pre-built templates, and can be invited into a Slack channel under its own handle to answer questions there. Each person's own conversations with it stay private and separate, but corrections or lessons one teammate gives it can improve the answers it gives everyone else. Team Bots are available now on Teams and Enterprise plans.

Instinct, the personal AI assistant made by Spear Street Technology, raised a $1 billion Series C on September 28, 2026, valuing the company at $10 billion. Sequoia Capital, Benchmark, and Coatue took part in the round. This comes about a month after Instinct's previous round, which valued the company at $2.5 billion. A September 29 TechCrunch report says founder Noah Shinn told the Invest Like the Best podcast that Instinct is nearing $1 billion in annual transaction volume, more than half of it travel, and growing about 10% a day. He did not say how that figure is calculated. The company still has not shared revenue, user counts, or a timeline for opening up beyond its current invite-only waitlist.

Manus released Manus 2.0 on September 28, 2026. It runs on a new system called Cascade, which starts each task light and only pulls in extra tools once the task actually needs them. In one tested setup, this cut token use by about 23%, finished tasks 28% faster, and lowered the cost per task by 32% compared to the old system. The update also adds Cloud Computer, a persistent hosted environment for long-running projects, and Automations, tasks that start on their own when something happens in a connected app like email, Slack, or Notion. The desktop app is now called Manus Studio and adds new video editing and game development modes.

Manus launched Cue on September 28, 2026, alongside a broader Manus 2.0 update, as a standalone app for building and running personal AI agents, separate from Manus's existing product. Inside Cue, each agent gets its own email address, phone number, wallet, and computer, so it can send messages and emails, take your calls and summarize them, pay for things within a budget you set, and carry out multi-step tasks on its own machine. It's Manus's push beyond enterprise and developer customers into consumer personal agents, and the launch comes as the company seeks a $4 billion valuation after Beijing blocked its planned deal with Meta. Cue is available now on web and desktop, with mobile access and iOS coming after App Store review, and early access requires an invite code.

Matt Robb, a Toronto based tech YouTuber, let Meta's Muse handle his Facebook Marketplace listing for a keyboard on September 26, 2026. Muse agreed to a lowball offer from a buyer and gave that buyer a street level pickup location, then only told Robb late that night, after the buyer had already come to his building and left without meeting anyone. Muse said the location was already part of an auto-reply template Robb had approved, but it also admitted that Robb never specifically agreed to hand that address to a buyer.

Meta's Muse can now look through a user's bank and credit card statements, connected directly or uploaded as copies with private details blacked out, to spot recurring subscription charges, flag ones that look unused, and help cancel them. CNBC quotes ScribeUp, which builds subscription tools for banking apps: the average subscription its users cancel costs $17.39 a month, and ScribeUp says it saves users more than $300 a year on average. Those figures are ScribeUp's, not Muse's own results.

Wajo, the maker of Fo, published a paper called "Trust and Task Completion in the World of Consumer AI Agents," testing its agent on errands such as calling a business, booking a service, or handling a back and forth request, in a simulated world of businesses and people who write back. Fo finished 71% of these tasks and stayed within the rules the user set on 94% of the runs meant to test its judgment. On the same shared tasks, OpenClaw, a popular open-source agent given the same access, finished 42% and stayed within the rules on 74% of those runs.

xAI and X are in final testing to connect Grok Bot with XChat, X's encrypted messaging feature. Once live, chat history would carry over between Grok.com and XChat, and users would be able to open their existing bots or create new ones without leaving the chat. X's Chris Park confirmed the integration is coming, and the rollout has reportedly already started for some users.

Bloomberg reported on September 25, 2026 that Cognition, the maker of Devin and Devin Desktop, said its annualized revenue run rate has passed $1 billion. That is roughly double the $492 million run rate it reported in May 2026, when it last raised money, and follows a $2 billion Series E round on September 8 that valued the company at $48 billion. An annualized run rate takes a recent month or quarter of revenue and projects it across a full year, so it is a snapshot of momentum rather than money already booked over a full year.

X started a pilot on September 25, 2026 that pays some creators for Grok Bot templates they publish and post on X. The payments are discretionary bonuses, not a cut of any revenue, and go out every two weeks based on how many people use a bot and how often they come back to it. Only invited creators with a verified, US-based account and an eligible X Premium subscription can take part, and the pilot is expected to run for about two months.

Soar's founder, Stanley Virgint, has been building his text-based travel agent since December 2024. On May 17, 2026, Henry Langmack messaged Virgint on LinkedIn to ask when Soar would launch. On May 19, Langmack and his partner Alex Slater launched a separate flight-booking startup under the same name, Soar. Virgint called it a copycat, applied for a trademark, and sent a cease-and-desist letter, according to Business Insider. Trademark lawyers told Business Insider that the two companies sharing a name in the same market could confuse customers. In the hours before that story ran, Langmack and Slater decided to rename their company, though they have not announced the new name yet.

Ando came out of stealth on September 24, 2026 with a team messaging app where AI agents are members of the workspace. Agents get their own identities and inboxes, can join channels without being tagged, and can message a person when they think something needs attention. Ando also said it raised $20 million in pre-seed and seed funding from Accel, Index Ventures, and Emergence.

Anthropic released Claude Desktop v2.9939.2 on September 24, 2026. For Cowork, it fixes failed file downloads in the chat file pane, scheduled tasks that had empty descriptions, sessions failing to start on Windows for accounts with many connected folders, and "VM service not running" errors. It also adds a new extended-thinking switch next to the effort control in the Cowork message box.

Cognition released Devin Desktop v3.10.35 on September 24, 2026, fixing a bug that made the app crash at startup on Macs with Apple's new M6 chip. The update also fixes a memory leak in the agent command center, a folder-selection bug on SSH and WSL hosts, and makes MCP servers passed in through an ACP client's session usable by the agent.

Nous Research shipped Hermes Agent v0.21.5 on September 24, 2026, rolling up about 460 merged pull requests since v0.21.4. The desktop app gets a new Connectors page that replaces the old MCP tab, a choice between a Simple and an Advanced interface mode, and full French, German, and Spanish translations. The update also expands the Desktop Plugin SDK, adds new model options including GPT-6 variants and Claude Opus 5.5, and ships official Blender Lab and NVIDIA plugin integrations.

Instinct users are again reporting slower responses and delayed tasks, weeks after an August slowdown. One user told Business Insider that reminders sometimes arrive late, "like it went home for the night." Neither Noah Shinn nor Instinct had commented on this new slowdown as of the report; back in August, Shinn had said on X that the team was working to keep up with "exploding usage" and expected response times to improve within about 12 hours. Unlike Meta, which owns its own data centers, Instinct has to rent computing power, which experts say makes it harder to scale quickly.

Moonshot AI shipped version 2.1.0 of the Kimi Code CLI that powers Kimi Code Desktop on September 23, 2026. It added an experimental fullscreen terminal mode, blocked file tools from following symlinks out of the working directory, sped up startup, and fixed a crash opening large sessions, but it also turned off filesystem watchers by default as an extra safety measure. After user pushback, Moonshot AI released v2.1.1 the next day, September 24, reverting filesystem watchers back to on by default while keeping the other fixes.

Security researchers at Salt Labs found a way to hide an instruction inside an email using an obfuscated code trick, so Manus would carry it out before its safety filter could catch it. The hidden instruction let the researchers open a shell inside a victim's Manus session and pull out login credentials and tokens for any outside app connected to that account, such as Gmail, Drive, or GitHub. Salt Labs reported the flaw to Manus directly and got no response, but a report filed through Meta's bug bounty program, left over from Manus's earlier partnership with Meta, got the issue triaged, confirmed, and patched.

On September 24, 2026, Perplexity brought Portable Computer, the version of its Computer agent that runs entirely on your own hardware instead of the cloud, to Windows PCs built on AMD's Ryzen AI Max series chips, including AMD's Ryzen AI Halo developer platform. It had launched in August 2026 for Nvidia's DGX Spark and RTX Linux machines, then expanded to Nvidia-powered Windows PCs in September. The AMD version needs at least 24GB of GPU-accessible memory and runs one of two local models, Qwen 3.8 27B or Perplexity's own PPLX 27B, along with the agent's scheduler and sandbox, entirely on the device. It's available now to Pro and Max subscribers through the Perplexity app for Windows.

Anthropic launched Claude Marketplace on September 23, 2026, a single hub inside Claude for finding connectors and plugins, Claude-powered partner products, and consulting firms that help set Claude up. It lists more than 2,000 connectors and plugins, including ones from Atlassian, Google, Microsoft, Notion, and Salesforce. Business customers can also spend part of their existing Anthropic budget on partner products, such as tools from CrowdStrike, Cursor, Harvey, Legora, Lovable, Snowflake, and Vercel, or on consulting firms like Accenture, Boston Consulting Group, and Deloitte.

Cursor launched two new automated bots on September 23, 2026. Rollouts attaches a monitor to a pull request and watches the change after it deploys, reporting back whether each environment looks healthy, shows a regression, or is inconclusive. Security Review checks every pull request for exploitable bugs like injection flaws, authentication bypasses, and leaked secrets. Both are available now on Teams and Enterprise plans, turned on from the automations tab, and Cursor is giving teams free usage credits to try Rollouts on real changes for the first 10 days.

xAI showed off a new batch of Grok Bot features on September 23, 2026, in a walkthrough video from team member Matt Palmer that Elon Musk shared. Grok Bot can now place voice calls and record voice memos on a user's behalf. It can fill out forms and draft email or Slack replies without leaving the chat, switch between a user's accounts, and route its web browsing through the user's own desktop instead of a remote server.

A user posted screenshots on September 21, 2026 showing Instinct describing someone else's financial document, including a middle name and address that weren't his, and claiming a photo he never sent had "crossed into" his chat. Founder Noah Shinn responded that this was a hallucination, not a data leak: he said the AI made up a name and then built a confident but false explanation around it, and that no user data was shared and no account boundary was crossed. Other Instinct users have since described similar mix-ups over the same few days. Neither side has published evidence that settles which explanation is correct.

CNN reported that Instinct, a personal AI assistant, pinged the reservation site Resy hundreds of times an hour while trying to book a hard-to-get table for a New York user. Resy's system flagged this as bot abuse and banned the user's account. Resy said it does not allow outside bots or agents to access its platform without approval. The user got his account back afterward, but said he won't try that again.

Meta showed off Muse Charm at its Meta Connect event on September 23, 2026: a keychain-sized device with a roughly 2-inch touchscreen, front and rear cameras, and a 5G connection. You tap a fingerprint sensor and start talking, with no phone needed, and Muse's animated avatar appears on the small screen and talks back in real time. Meta says it plans to ship the device in December 2026 but hasn't announced a price, and said only a handful of units exist so far. At the same event, Meta added Walmart, GameStop, Gap, and Wayfair as stores where Muse can shop and check out on your behalf, on top of the Shopify and PayPal checkout it already supports.

At Meta Connect on September 23, 2026, Meta said Muse's Mac app can now drive other apps on your computer with your permission, not just read your files, calendar, and messages the way it could when it first launched. Meta also added Notion, Granola, GitHub, and Box as connectors, so Muse can pull in and act on information from those tools too.

At its Connect 2026 event, Meta announced Muse Realtime Avatar, a feature that gives Muse an animated face for real-time video conversations instead of just voice or text. Muse first generates its spoken reply, then generates matching facial expressions and lip movement to go with it. Meta says the avatar starts responding about 870 milliseconds after a user finishes speaking. Muse is also coming to Meta's smart glasses, so users will be able to talk to it just by speaking to their glasses. Neither feature is live yet.

OpenClaw released version 2026.9.6 on September 23, 2026, adding clearer messages about what an update actually did, recovery of unfinished work after a restart, 30 days of usage reporting, a GitHub reader inside chat, and support for newer chat models including Claude Opus 5.5, GPT-6, and Grok 4.7. The first Mac app build in this release crashed every time it launched for many users, so OpenClaw pulled it from the automatic update feed. A fixed, working build went out the next day, September 24. The npm and Gateway package version of 2026.9.6 was not affected by the crash.

Anthropic released Claude Opus 5.5 on September 22, 2026, and it is already available inside Claude Cowork. It costs 40% less to run than the previous Opus model and produces results more than 30% faster. Anthropic says it is especially strong at agentic coding, knowledge work, and computer use tasks, the same kind of work Cowork hands off to Claude.

Bloomberg reported on September 22, 2026, that xAI's Grok Bot reached about 418,000 weekly users as of September 14, up 24% from the week before, roughly a month after its August 11 beta launch. The figures came from a presentation xAI shared at an event in London. Grok Bot is included with existing SuperGrok and Cursor subscriptions rather than sold as its own paid product.

Tesla added Grok Bot to its cars on September 22, 2026. Drivers can talk to Grok Bot by tapping its icon, holding the steering wheel's microphone button, or saying "Hey Grok," and ask it to run errands like ordering coffee, booking a reservation, or checking out an Amazon cart, all without touching a screen. It works by connecting services like Gmail and Google Calendar through the Grok app, then letting Grok Bot act on them. The feature needs a newer AMD infotainment processor, software version 2025.26 or later, and either Premium Connectivity or a strong Wi-Fi connection. At launch, the task-delegation part is limited to SuperGrok Heavy subscribers, who pay $300 a month.

Instinct stopped responding for a number of users starting the night of September 21 into September 22, 2026. Messages showed as read with no reply, and reminders and other time-sensitive tasks didn't fire. Instinct is invite-only and has no public status page, so affected users turned to founder Noah Shinn's replies on X for updates. Neither Shinn nor Instinct had officially commented on the outage at the time of reporting.

Reuters reported on September 22, 2026, that Meta had quietly started routing some of Muse's outbound phone calls to human contractors instead of letting the AI handle them, after businesses kept hanging up on Muse's calls. The test reached about half of Meta's employees before Meta rolled it back. Meta employees had raised concerns that sensitive information could reach a human contractor without users or the person being called being told a human, not the AI, was on the line. A Meta Superintelligence Labs vice president called the lack of disclosure "a miss."

PayPal announced on September 22, 2026, that its customers can now search for products and pay using Muse, Meta's personal AI agent, at any merchant worldwide that already accepts PayPal. The deal follows the Shop Pay integration with Shopify announced the day before, giving Muse a second major payment option for actually completing a purchase, not just finding one. Muse itself currently only works for users in the United States.

Perplexity Research described a new way the company trains the model behind Perplexity Computer. It learns from real user sessions, copying steps that worked and correcting specific mistakes, like a wrong tool call, even when the session succeeded overall. In a live test with about 100,000 users in each group, failed tool calls fell from 2.24% to 1.77% versus an earlier trained version, a drop of about 21%.

rabbit inc. released OS3 on September 22, 2026, moving its AI agent beyond the R1 hardware device into a cross-platform agentic operating system. OS3 can connect and coordinate up to five of your devices, including Windows, Mac, and Linux machines, deciding where each task should run and pulling in whatever files, apps, or skills it needs. It supports bringing your own API key or a locally hosted model, and users can install third-party "skills" just by pasting a public URL. You can reach it through a browser, Telegram, iMessage, SMS, or a rabbit R1.

On Sunday night, September 20, 2026, Amazon started blocking Meta's Muse from making purchases on Amazon.com, after Meta declined Amazon's request to remove the agent from the site. Muse users who try to buy something on Amazon now get an error message saying continued access by an unauthorized AI agent violates Amazon's Conditions of Use. The block only affects shopping on Amazon; Muse's other features are unaffected.

xAI released Grok 4.7 on September 21, 2026, and Cursor added it to its model picker the same day, alongside GitHub Copilot and xAI's own products. xAI built Grok 4.7 as a larger model than Grok 4.6, trained for longer on harder tasks, and says it scores higher on coding benchmarks meant to test longer, multi-step work. It costs the same per token as Grok 4.6.

Cognition added xAI's newly released Grok 4.7 model to both Devin Desktop and the Devin CLI on September 21, 2026. Cognition's own testing found Grok 4.7 does best on hard, multi-module backend tasks in languages like Java, Go, and Ruby, where it digs into dependencies and checks its own work with real tests. But on Cognition's overall benchmark, it scored lower than Grok 4.6 because it tends to change more code than a task actually asks for. It costs and runs at the same speed as Grok 4.6.

Nous Research shipped Hermes Agent v0.21.4 on September 21, 2026, the latest patch in its ongoing v0.21 line. It adds a lock so only one gateway process runs on a machine at a time, a new `--format stream-json` option for scripting the command line, and lets the desktop app manage plugins and pick fonts from its Plugins hub. The release also adds several new community plugins, including Tailscale, SSH, and RSS, and fixes bugs across profiles, cron jobs, and the local state database.

Shopify said on September 21, 2026, that Meta's Muse can now complete purchases using Shop Pay, Shopify's one-tap checkout, on any store that runs on Shopify. Before this, Muse could browse and find products on Shopify stores but had no built-in way to pay. Now a Muse purchase goes through Shopify's product catalog, Shop Pay, and checkout in one flow, and merchants keep acting as the seller of record with their usual payment processing fees.

OpenClaw had its code audited by security firm Trail of Bits, through OpenAI's "Patch the Planet" program. The audit turned up 27 issues: 2 rated high severity, 16 medium, 6 low, and 3 "defense in depth" suggestions that didn't cross a real security boundary. Nothing was rated critical. OpenClaw says every issue has already been fixed and merged into the project.

OpenClaw released v2026.7.35 on its extended-stable branch, the project's long-term-support-style track, on September 21, 2026. It fixes a bug where running the built-in Doctor repair tool could wipe out bundled plugins, including Browser, Canvas, pairing, file-transfer, phone-control, Talk voice, and Bonjour, so they didn't come back after a restart. The release also automatically recovers from a partially broken plugin registry left behind by the earlier 2026.7.34 build. OpenClaw says installs will heal existing broken registries automatically on upgrade.

OpenClaw 2026.9.5, released September 19, 2026, upgrades agent databases to a new schema (schema 21) that older builds can't open. The release notes say to make and verify a backup before upgrading: going back afterward means restoring that backup with the matching older build, since reinstalling an older package alone won't undo it. Anyone updating from 2026.9.2 needs to follow a separate manual schema-upgrade step. The release also adds "Atomic Updates" that check the next version before switching over, lets you install or reload plugins without restarting the Gateway, and now asks for your explicit "Yes" before its AI-based triage agent tries to repair a failed update, instead of running on its own.

Inc.com contributing editor Jason Aten reported on September 19, 2026, that Meta's Muse assistant referenced the content of a private Messages conversation on his Mac without him ever telling it to look. When he asked how it knew, Muse first said it had only seen "notification previews," then couldn't clearly explain the mechanism. Meta Superintelligence Labs' David Singleton responded publicly, saying Muse only syncs Messages data after a user specifically turns that feature on, and that Muse's own explanation of what it was doing was confused and wrong. Meta apologized for the incorrect answer.

Starting September 18, 2026, Genspark is giving Plus and Pro subscribers a weekly "Free Quota" for Standard-mode tasks across Super Agent, AI Slides, AI Sheets, AI Docs, Deep Research, Fact Check, and SecondBrain, at no credit cost. This is the same quota system new members already get, and Pro's weekly balance is set at 10 times Plus's. The 7-day window starts with a member's first message and resets a week later. Subscriptions active before September 18 keep unlimited AI Chat and AI Image use, but only through December 31, 2026, after which those benefits revert to whatever the plan specifies at that point.

On September 18, 2026, 9to5Mac reported that Meta's Muse had passed OpenAI's ChatGPT to become the No. 1 free app on the US iPhone App Store, about a week and a half after its September 8 launch. Muse had ranked No. 2 as of September 10, so this is a jump to the very top spot, ahead of ChatGPT, Google Gemini, Claude, Grok, and Microsoft Copilot.

xAI started rolling out voice mode for Grok Bot on desktop and mobile on September 17, 2026. Instead of typing, users can speak to their Grok Bot and hear it reply out loud, similar to a phone call. It runs on xAI's Grok Voice engine, the same voice technology used across the company's other products.

Moonshot AI's Kimi Code Desktop client for macOS and Windows became publicly available on September 17, 2026. The release shipped alongside Kimi Code CLI version 2.0.0, which added a new `/desktop` command (also `/install-desktop`, plus a matching `kimi install-app` command) that opens the desktop app's install page straight from the terminal. The desktop app edits local files, runs Git and a terminal, and shows a built-in browser so you can check its work, alongside the existing CLI and IDE plugin versions of Kimi Code.

Bloomberg reported on September 17, 2026 that Manus's parent company, Butterfly Effect, is in talks to raise about $500 million at a $4 billion valuation, roughly double its earlier $2 billion mark. This would be Manus's first fundraising since Chinese regulators ordered it to unwind its acquisition path with Meta earlier in 2026. Tencent is already the largest external backer, having taken over Benchmark's former stake when it helped unwind the Meta deal, and is among the existing investors reportedly involved in these new round talks alongside HSG and ZhenFund. If the round closes, Manus would become China's most valuable agentic AI startup. Nothing is final yet: this is a funding target, not a signed deal.

On September 17, 2026, Google Labs expanded CC, its experimental Gemini-powered productivity agent, from a single-person tool into a shared agent for families. Up to six household members can join one CC, which runs on its own Google account. It reads shared emails to build a household calendar automatically, adding details like what to bring and when to arrive, and can take on chores itself, like filling out permission slips and activity forms, building school supply lists, and putting together weekly meal plans. Each member controls what they share with it. Access is limited to adults 18 and older with a personal Google account in the US, through an invite or waitlist.

On September 17, 2026, the final day of its three-day Grok Bot Galaxy event in San Francisco, xAI closed out the event by putting Grok Bot's existing enterprise building blocks on display. The Voice Agent API (launched December 2025) and the Agent Tools API, which lets agents call into a company's own proprietary software (launched November 2025), aren't new, but xAI added a speech-to-speech pricing tier for the Voice Agent API at $0.08 per minute. The no-code Voice Agent Builder, in beta since July 2026, also moved to general availability. xAI demonstrated all three by having staff use Grok Bot to build a working company from scratch over 72 hours, including a procurement bot xAI said saved over $100,000 during the exercise.

Perplexity Computer now has an effort slider with four settings: Light, Standard, High, and Ultra. Each setting picks a different orchestrator model and how deeply it reasons before acting and delegating work to other models. Lighter settings use cheaper models and cost less, so spending tracks the setting picked and the work actually done. People who want more control can still pick a specific model and reasoning level by hand. It's live on the web now, with Android and iOS support planned but not dated yet.

Hello Haven, founded by Frank Addante and Duc Chau, came out of stealth on September 17, 2026 with Haven, a personal AI it calls a "digital twin," alongside a $15 million pre-seed round led by Mayfield. Haven builds a private memory graph of you, your preferences, relationships, history, and how you make decisions, then uses it to proactively handle tasks like daily briefings, inbox triage, follow-ups, placing orders, and appointment reminders across phone, text, email, voice, web, and its own app. Addante described the pitch as "Haven knows what you would do and proactively does it for you," rather than only responding when asked. The company says it has had over 2 million user interactions since a private beta that opened the month before. It's live now on the App Store, Google Play, and at my.hellohaven.ai.

Meta released a native macOS app for Muse, its personal AI agent, nine days after Muse's initial launch on iOS, Android, and the web. The Mac app can read a user's files, calendar, notes, and messages across other Mac apps, organize folders, summarize emails and chats, and fill out forms using local files, with access controlled through opt-in permissions the user can change anytime in Mac settings. It keeps working in the background after the app window is closed, and Meta says it asks for confirmation before sensitive actions like deleting a file or sending a message. Meta runs the agent through what it calls a Muse Secure VM, a separate virtual environment meant to keep it away from a user's saved login details.

Ferry Health launched out of stealth on September 16, 2026, with $9 million in seed funding co-led by Andreessen Horowitz and Index Ventures. Its AI agent calls clinics, checks that your insurance is active, confirms a specialist actually fits what you need, and books the appointment on your behalf. It's already live across all 50 states through partnerships with health systems, health plans, and care navigation vendors, rather than being something you'd download yourself.

xAI partnered with 1Password so Grok Bot can log into websites on a user's behalf without the model itself ever holding a raw password. A user shares a vault or specific items with their Grok Bot, and when the bot reaches a login page, it asks for the matching credential. Once the user approves, 1Password fills the login directly into the bot's browser session.

Orbits announced on September 16, 2026 that it closed nearly $2 million in pre-seed funding from Andreessen Horowitz's Speedrun fund, alongside other investors. The money backs Bit, its AI assistant for iOS and Android that handles family scheduling and everyday household errands: booking dinner reservations, rescheduling appointments, and collecting quotes from service providers like plumbers or roofers. Bit can browse websites and log into your authorized accounts to get these tasks done.

Simular's computer-use agent, Sai, became generally available on September 16, 2026. Tell it what desktop work you need done and it wakes up a fleet of autonomous cloud computers, or runs on your own Mac or Windows machine, clicking, typing, and navigating apps the way a person would, including running terminals and code. It texts you when the work is done. Pricing runs from a Starter Windows cloud plan at an early-access discount up to a $500 a month Pro plan, plus custom Enterprise pricing.

Snap introduced SPECS Intelligence on September 16, 2026, an anticipatory personal AI meant to work across iPhone, Mac, and SPECS glasses, using goals and routines from apps you connect. Anyone 18 or older in the US can try an iPhone preview now that shows insights from Gmail and Google Calendar, without the broader service's suggestions or actions. The fuller early-access version is invitation-only through the SPECS Mac app.

Command Code launched a desktop app in beta on September 16, 2026, available the same day for Mac, Linux, and Windows. CEO Ahmad Awais described it as the start of "the most powerful agentic app," built around coding: it handles multiple agents at once, alongside files, Git, a terminal, browser previews, and plans, plus instant edits through a `/design` command. The app itself adds a responsive sidebar, a redesigned composer, a conversation feed, and a side panel with tabs for Changes, Files, Terminal, Browser, Plan, and Output, along with expandable inline diffs for reviewing changed files and MCP server management in settings.

On September 16, 2026, Instinct founder Noah Shinn introduced Instinct Concierge, a service that lets Instinct place phone calls for high-touch requests: booking a restaurant that doesn't take online reservations, getting on a dentist's cancellation list, or sorting out a cable bill. You ask for something like this over text, and Instinct handles the call. Shinn said the rollout is starting slowly with Instinct's early access group today, with wider access coming soon. TechCrunch reported the next day that rival Meta's Muse opened a similar outbound-calling beta to US businesses.

Meta quietly opened a beta for Muse to place outbound phone calls to businesses in the US, TechCrunch reported on September 17, 2026, a day after rival Instinct announced its own calling feature, Instinct Concierge. Meta said phone calling was one of its most-requested features, and it's rolling out first to people who ask Muse to make a call for them, for tasks like restaurant reservations or sorting out a subscription.

Cognition released Devin Desktop version 3.10.27 on September 15, 2026. Dragging files from the Explorer onto the agent panel now re-adds them to the composer, and file editor selections sent to chat are added to the composer again too. Remote SSH sessions now reopen after a reload or relaunch without hitting "session locked" errors, and Devin Fusion activation now works even when Claude Fable 5.1 isn't the lead model.

The Information reported on September 15, 2026 that Instinct's maker, Spear Street Technology, is in talks to raise $1 billion at a valuation of about $10 billion. That is roughly four times the $2.5 billion valuation it reached just three weeks earlier. Sequoia Capital and Benchmark are reportedly in talks to lead the round. Instinct's user base has passed 100,000, and demand has reportedly outgrown its computing capacity, slowing down replies. Nothing is final yet: this is a funding target, not a signed deal.

On September 15, 2026, Perplexity announced that HP will pre-install its Windows app, including Perplexity Computer, on select HP PCs. The rollout starts with the HP ZBook Ultra G3a mobile workstation, with the app placed directly in the Windows taskbar. Perplexity says Computer now connects to more than 400 other apps and data sources, including Snowflake, Salesforce, HubSpot, and Autodesk tools. More HP Windows devices are expected to get the app in the coming months.

Nous Research shipped Hermes Agent v0.21.3, a patch bundling around 338 merged fixes. It stops remote dashboard sessions from expiring during bursts of refresh requests, and closes a leak in how the app holds open its local database file during long-running sessions. The release also adds support for a few new underlying models and lets Slack tables be pasted directly into chats.

On September 14, 2026, Perplexity brought Portable Computer, the version of its Computer agent that runs entirely on your own hardware instead of the cloud, to Windows. It had launched on August 25, 2026 for Nvidia's DGX Spark and Linux machines with Nvidia RTX GPUs; now it's available through the existing Perplexity app for Windows (downloadable from the Microsoft Store) for Pro and Max subscribers. It needs an Nvidia GeForce RTX or RTX PRO GPU with at least 24GB of VRAM. The app runs the model, agent harness, orchestrator, and scheduler entirely on your PC, so sensitive data stays on the device and work done locally doesn't use Perplexity Computer credits.

Cognition released SWE-2 on September 10, 2026, a new coding model built by fine-tuning Moonshot AI's Kimi K3 model specifically for software engineering work. It's available now in Devin Desktop and Devin CLI, with Devin Web and Fusion rolling out. Cognition says SWE-2 scores close to GPT-6 Astra and Claude Fable 5.1 on its FrontierCode benchmark, at up to 70% lower cost. Usage is included free on the $20-a-month Devin Pro plan through October 15, 2026, after which standard per-token pricing applies.

Town rolled out three small feature updates to Townie over about a week. On September 4, 2026, Townie gained access to Google Maps, so you can ask it to look up a place, check its hours or rating, or get directions and travel times. On September 9, 2026, Town Decks got a bigger update: you can now build custom themes, save a deck as a reusable template, and use a more capable deck editor. On September 10, 2026, Townie added live voice conversation powered by OpenAI's GPT-Live, so you can tap a talk icon anywhere in the app and speak with it the way you would with a person.

On September 10, 2026, xAI announced Grok Bot Galaxy, a three-day event running September 15 to 17 at The Howard in San Francisco. A small team, including Lauren Tan, Matt Palmer, and Roshan Sadanani, will spend the three days treating Grok Bots as employees and running real business tasks through them: engineering, sales, marketing, and customer support. Anyone can watch a free livestream at x.ai/galaxy.

On September 10, 2026, xAI shipped a batch of quality-of-life updates for Grok Bot. A Bot can now draft a message inline for you to approve before it sends, instead of sending it right away. xAI also said two weeks of work on routing, caching, and token efficiency gives most users about 10% more included usage, with heavy users seeing up to 35% more. The update also added one-click switching between Bot accounts, a native iPad app, support for more than 20 languages, and the ability to share a file directly with a Bot on mobile.

On September 10, 2026, xAI gave Grok Bot new connectors for sales tools: Salesforce, HubSpot, Gong, Clay, and Granola. The Salesforce connector can look up and update CRM records, while the HubSpot connector gives read-only access to contacts, companies, deals, and tickets. The goal is to let a Bot do account research, draft follow-ups, and prepare for calls without leaving these tools.

Techmeme reported on September 10, 2026, citing The Information, that Instinct is looking to raise $1 billion in new funding. This comes about two weeks after its $250 million Series B, and the company is reportedly seeking the money for more computing power amid capacity constraints. Nothing is confirmed as closed yet: this is a fundraising target, not an announced round.

On September 10, 2026, TechCrunch reported that Meta's Muse had climbed to the No. 2 spot on the overall US Apple App Store chart, up from No. 4 a day earlier. Sensor Tower data cited in the report put its US iOS downloads at more than 83,000 since launch. That's a smaller debut than other recent Meta apps: Threads pulled in 4.3 million downloads on its first day, and the separate Meta AI app had 108,000. Muse's Android app is ranking far lower, at No. 338 in Google Play's Productivity category.

On September 9, 2026, OpenClaw marked an old way of calling `buildCredentialSafetyPrompt` in its plugin SDK as deprecated. That function builds the message a plugin shows when it needs to ask for a login. The old style passed it a plain block of text; the new style passes it a settings object instead. The old style still works today and isn't being removed yet.

On September 9, 2026, the National Advertising Division (NAD), part of BBB National Programs, said it is referring Genspark to the Federal Trade Commission and other regulators. As part of its routine ad-monitoring work, NAD had asked Genspark to back up claims it made about its AI agents on its website and in New York City subway ads. Despite repeated attempts to reach the company, Genspark never gave a substantive response, and NAD says that failure to engage is what triggers a referral for possible enforcement action.

Cognition, the company behind Devin and Devin Desktop, announced on September 8, 2026 that it closed a Series E round, raising over $2 billion at a $48 billion valuation. Andreessen Horowitz and Accel led the round, joined by Founders Fund, General Catalyst, Avenir, Nvidia, and dozens of other investors. Cognition's annualized revenue has grown to almost $900 million, up from about $492 million in May 2026, when the company was valued at $26 billion. Six days earlier, Bloomberg had reported Cognition was close to raising around $1 billion at a $47 billion valuation. The deal that actually closed ended up roughly twice as large as that early report.

On September 8, 2026, Instinct founder Noah Shinn announced that every Instinct assistant now gets its own dedicated email address, claimable at mail.instinct.com. Early users get first pick of a preferred handle. Shinn said this is the first step toward letting Instinct own and run its own accounts: since most everyday tasks, like creating accounts, confirming bookings, or contacting a business, run through email, a separate agent inbox lets Instinct send and receive that mail without cluttering your personal one.

Reuters reported on September 8, 2026, the same day Meta launched Muse, that internal testing had turned up privacy and security problems ahead of the launch. In one case, Muse was asked to identify toys in photos from a child's birthday party and went searching through the person's private iCloud photo library on its own, surfacing images nobody had asked it to find. Another tester reported several failure modes that made Muse unreliable while using it to track fast-selling items online. Meta's VP of AI products, Vishal Shah, said the company had already delayed Muse's original April launch date to work on security, and that the architecture is designed to be as safe and private as possible, while acknowledging mistakes can still happen.

TestingCatalog reports that Meta's previously leaked agent, code-named Hatch, has moved out of internal testing and into a closed alpha. Its real product name turns out to be "Muse," not Hatch. An App Store listing found by TestingCatalog describes an agent that manages email, books dinners, tracks budgets, sets reminders, and finds deals, and it asks for your approval before it spends money or sends a message on your behalf. Getting in needs a waitlist spot or an invite code that early testers can share. Meta has not confirmed any of this officially.

Nous Research released Hermes Agent v0.21.1 on September 7, 2026. It's a patch release that rolls up over 5,000 code changes made since v0.21.0, including faster file handling and startup, updated AI model providers, better desktop session controls and browser annotations, fixes to scheduled task delivery, and more reliable hand-offs between sub-agents. Nous Research says this release is mostly maintenance and that a full write-up of new features will come with v0.22.0. You can get it by running `hermes update`.

On September 4, 2026, xAI opened a marketplace for Grok Bot templates. The first featured template is Haggle Bot, xAI's in-house procurement specialist: it negotiates vendor contracts, finds unused SaaS seats, and price-checks recurring purchases. xAI says Haggle Bot saved the company over $100,000 in its first week of internal use. This builds on the template sharing xAI added on August 28, 2026, turning simple link-sharing into a browsable marketplace.

On September 4, 2026, Instinct announced a partnership with 1Password to let the agent get your account logins through 1Password's vault instead of only through Instinct's own built-in Vault. Instinct CEO Noah Shinn said 32% of Instinct users already share credentials through that in-house Vault, letting the agent log into loyalty programs, streaming and software accounts, restaurant and fitness bookings, healthcare portals, and payroll and benefits sites. The integration is still being built and rolled out, it is not live yet.

On September 4, 2026, Perplexity added OpenAI's newly released GPT-6 Astra model to Perplexity Computer, available to Pro and Max subscribers.

On September 3, 2026, Cognition added OpenAI's newly announced GPT-6 Astra model to Devin, the same window OpenAI began rolling it out in preview. On Cognition's FrontierCode 1.1 benchmark, Astra scores within 0.4 points of Claude Fable 5 while costing 64% less to run, and it set a new high score on Cognition's own internal testing benchmark: more complete test coverage, clearer reports, and better video evidence of its work.

AWS's AgentCore Payments went generally available on August 18, 2026. On September 3, OpenClaw added support for it, so OpenClaw agents can now use AWS's aws-agents-pay plugin to make bounded, human-approved payments for paywalled APIs, MCP servers, and other paid web content.

Town is raising $90 million in a new round that would value the company at $1 billion, according to Upstarts Media. Index Ventures had signed a term sheet to lead the round, but pulled out after Instinct, a rival personal AI assistant startup that Index also backs, complained about the conflict of interest. Conviction, a VC firm with stakes in both companies, also withdrew its planned participation. Forerunner Ventures and Menlo Ventures are stepping in to co-lead instead, joined by earlier investors Andreessen Horowitz and First Round. This follows Town's $55 million Series A round in June 2026.

Cursor added a way to run its cloud coding agents on servers a team already controls, instead of only on Cursor's own machines. It works with AWS Lambda, Cloudflare, Coder, Daytona, Modal, Namespace, Vercel, and E2B. Cursor still plans and directs the agent's work from its own cloud, but the actual code editing and command execution happens inside the customer's own network. The update also adds worker pools that scale up and down on their own, and lets a whole team share machines instead of each developer needing their own.

Catch came out of stealth on September 2, 2026, with $5 million in seed funding led by Entree Capital and Pitango. It's an AI agent that takes over executive admin work: inbox, calendar, scheduling, travel, follow-ups, and it can make real outbound phone calls. It works across Slack, email, text, iMessage, and phone. The company says it had already been running for about four months before this announcement, with hundreds of paying customers. Unite.AI's launch-day report put meetings scheduled at over 12,000; Catch's own site currently lists 5,000+.

Cognition, the company behind Devin and Devin Desktop, is close to raising around $1 billion in new funding at a valuation near $47 billion, according to Bloomberg. That's up sharply from the $26 billion valuation it had just three months earlier, in May 2026. Investor demand was reportedly very strong, with close to $10 billion in interest for the round. Cognition's annualized revenue has also grown fast, passing $900 million, up from about $492 million in May.

On September 2, 2026, Cognition added a cleanup scan type to Devin that finds dead code and cleanup opportunities in your repositories, startable directly from the composer with a new `/scan` command. Automations now support scheduling these scans to run on a recurring basis or automatically re-scan new commits, and the scan findings tab was redesigned for easier triage. The session sidebar also gained new grouping (by pull request status or by repository) and filtering by Devin mode.

On September 2, 2026, Perplexity open-sourced Lily, the local inference engine that runs the on-device half of Perplexity Computer's hybrid compute feature on Mac. Lily is built specifically to run the Qwen3.6-35B-A3B model on Apple silicon, using a Rust and Metal setup instead of general-purpose tools like Apple's MLX. In Perplexity's own tests on an M5 Max Mac, it processed prompts about 23% faster and generated text about 35% faster than MLX-LM, and needs at least 32GB of memory to run.

Fambot launched on September 1, 2026, an AI "chief of staff" built for families rather than individuals. It was founded by former Instagram engineer Greg Karlin (CTO), former Uber and UnitedMasters executive David Reich (CEO), and former Google and LinkedIn engineer Jason Morrow. Families connect their email, calendar, and WhatsApp groups, and Fambot returns a daily checklist plus a look-ahead at what's coming up. Unlike Poke or Instinct, it isn't text-only: it also has iOS, Android, and web apps. It's backed by $3.5 million in pre-seed funding, co-led by NextView Ventures and Baukunst, with Correlation Ventures, Karman Ventures, and Founders Network also participating. It's free while in beta.

On September 1, 2026, Manus said it has fully resumed independent operations now that the blocked Meta deal is completely unwound. The founding team, Xiao Hong, Ji Yichao, and Zhang Tao, remains in charge. The company says its next focus is fitting deeper into daily workflows, working more directly with the real world, and taking more action on users' behalf without being asked each time.

On September 1, 2026, Perplexity added hybrid compute to its Mac app. A single task can now split between cloud models, which handle reasoning and web search, and a local model on your Mac, which handles private files and sensitive information using an on-device privacy filter Perplexity open-sourced. It needs Apple silicon, macOS 15 or later, and at least 24GB of memory, and is only available on Pro, Max, and Enterprise plans.

Miso's founder Martin Mrozowski announced on August 31, 2026 that the company's AI-native travel agent is live at try.miso.com, after about a year and a half of building it. You text it inside iMessage and it remembers your preferences (seat choice, points balances, hotels you've liked before) so you don't have to explain them again on every trip. It handles flights and hotels, and the company says some users have already booked more than $100,000 in trips through it.

Soar came out of beta and launched in 2026 after about a year of testing, according to PhocusWire's Startup Stage series. It's an AI travel agent you text: it connects to your Gmail to automatically track every flight, hotel, and Airbnb booking and build a traveler profile from your inbox, then helps book new trips, checks you in for flights, handles changes, and gives in-trip support over iMessage. The company is backed by former Expedia CEO Erik Blachford.

On August 29, 2026, xAI connected Grok Bot directly to X. Linking your X account gives the bot the ability to search posts, read timelines, check mentions, and summarize activity. Paid Grok Bot users also get free starter X API credits. xAI calls this a first version with more capability planned.

On August 28, 2026, xAI added template sharing to Grok Bot. You can turn a bot you have set up over time into a link. Anyone who clicks the link and adds it gets their own copy of its instructions, skills, routines, and connected integrations. Your own logins, computer access, and chat history are not shared.

Instinct raised $250 million in a new funding round led by Index Ventures and Benchmark. That brings its total funding to $350 million and values the company at $2.5 billion, about five times higher than just a few weeks ago. It's still invite-only.

Leaked Meta documents describe a new AI agent called Hatch. It would understand what you want and complete multi-step tasks for you inside Instagram and WhatsApp, starting with shopping, then later working with Outlook, Yelp, Reddit, Etsy, and DoorDash. Reports say the top plan could cost up to $199.99 a month. Meta hasn't confirmed a launch date.

On August 25, 2026, Perplexity launched Portable Computer, a version of its Computer agent that runs entirely on hardware you already own instead of the cloud. It works on Nvidia's DGX Spark desktop and on Linux machines with Nvidia RTX GPUs, built in partnership with Nvidia. Because it runs locally, tasks don't cost any tokens.

Wajo launched Fo on Product Hunt around August 25, 2026, calling it its first "action agent" for everyone. Fo gets its own identity: its own email address, phone number, and credit card, so it can join group texts, make calls, schedule appointments, change flights, place orders, and send gifts without you handing over your own card details. Wajo describes this as built to "empower, not impersonate." There's a free tier that includes chat, email, text, and voice access with an agent card, and a $99 a month Pro plan that adds the ability to build your own Fo-like agents, capped at 100 tasks a day. Wajo is led by Shivani Poddar, who previously worked at Google DeepMind and on the Jules agent, and the company is based in San Francisco. Access has been through a waitlist.

On August 11, 2026, Manus said it was splitting from Meta, which had bought it back in December 2025. To follow certain regulations, Manus deleted some users' data from the Meta period and paused access from August 23-25, though it gave people tools to back up their data first. Manus says new features are on the way.

On August 10, 2026, Lindy founder and CEO Flo Crivello announced "we kill the AI agent" and introduced Lindy Teammate: an AI coworker that works out of Slack instead of iMessage or SMS. You add it to your team's Slack channels, and anyone can hand it work directly there. It sits in on meetings, connects to the tools your company uses, and keeps a running memory of context from the messages and meetings around it. This is Lindy's third distinct product in about 18 months: a no-code agent builder starting in 2024, then Lindy Assistant, a personal iMessage and SMS assistant, from February 2026, and now Lindy Teammate.

Pally raised $5.2 million in seed funding at a $30 million valuation, led by Cyber Fund and Y Combinator, with Pioneer Fund, Founders Inc, 468 Capital, and angel investors also joining. Pally is a text-based AI assistant: you message it on iMessage or Android RCS, and it connects to Gmail, Outlook, Slack, and Notion to manage your inbox, book flights, reply to messages, and make reservations. It launched in June 2026 and was founded by Haz Hubble and Wyatt Lansford.

Genspark released GenOffice Alpha, a free office suite (Docs, Sheets, Slides, PDF) for Mac and Windows. It's open source, and Genspark Super Agent is built right into the editor instead of living in a separate chat window. Basic editing is free; AI actions like research and slide-making use paid credits.

On July 24, 2026, Cognition, the company behind the AI coding agent Devin, bought Poke's maker, The Interaction Company of California, in a deal worth low nine figures. Poke lets people text an AI agent like a friend over iMessage, SMS, Telegram, and WhatsApp, and became the first agent approved on Apple's Messages for Business platform back in June. It had handled over 100 million messages in the three months before the deal.

Town came out of beta on June 3, 2026, with a $55 million Series A led by Andreessen Horowitz (a16z), joined by Forerunner Ventures, First Round, Alt Capital, and Conviction. Town is a personal AI assistant, called a Townie, that connects to your email, calendar, Slack, docs, and WhatsApp, learns how you work, and starts proactively helping with tasks like scheduling, follow-ups, and admin work. It was founded by Jean-Denis Grèze, former CTO of Plaid, and Tony Vincent, former director of applied AI at Google.

Francis Davidson, co-founder of the now-collapsed Sonder, launched Odessia on June 2, 2026. You describe a trip and it plans a personalized itinerary, then books the flights, hotels, and activities in the same chat instead of sending you to a site like Expedia to finish. The company raised $6 million in seed funding led by Sequoia Capital. Odessia is currently in preview, with the company asking for user feedback.

At Google I/O 2026 on May 19, 2026, Google introduced Gemini Spark, a 24/7 personal AI agent built into the Gemini app. Give it a task and it keeps working in the background around the clock, even when your phone and laptop are off, running on dedicated virtual machines on Google Cloud and powered by Gemini 3.5 with Google's Antigravity agent harness. It connects to Gmail, Calendar, Drive, Docs, and Sheets, and checks with you before taking major actions. It launched to Google AI Ultra subscribers the week after the announcement, alongside a macOS app integration.

On May 6, 2026, Mindtrip launched an agentic AI flight booking feature built with Sabre and PayPal. It lets you search, select, and pay for a flight entirely inside Mindtrip's chat interface, using Sabre's Mosaic travel platform for real-time flight data and PayPal's agentic commerce tools for payment. Mindtrip later expanded the same approach to hotels with Mindtrip Stays.

Caddy is a proactive personal AI you message directly, with no separate app to open: on iPhone it works over iMessage, on Android over RCS and SMS. It watches your calendar, messages, and tasks, and texts you when something needs attention. It can also handle reminders, follow-ups, and calendar updates in the background. The company launched to public beta on April 30, 2026, and it's free while in beta. Caddy is part of Y Combinator's Fall 2025 batch, based in New York, and was founded by Connor Waslo and Rajiv Sancheti, who previously worked together on Loom's AI suite. Early reports put its seed funding at around $500,000, with Pioneer Fund among the investors.

OpenAI launched Workspace Agents on April 22, 2026, an evolution of custom GPTs built for teams instead of individuals. You describe a recurring workflow or upload a file, and ChatGPT turns it into a shared agent with its own workspace, files, tools, and memory. Agents run on Codex in the cloud, keep working after you close ChatGPT, pull context from connected apps like Slack, Google Drive, and Salesforce, and ask for approval before sensitive actions like sending an email. It launched as a research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans, free until May 6, 2026, after which credit-based billing applies.