Between May and September 2026, OpenAI, Google, Meta, xAI, Anthropic and Microsoft all shipped the same idea: an AI that keeps working after you close the window. It has a name. It remembers you. It runs on a computer of its own, works through the night, and asks before it pays or sends anything.
On the scale the Octonomous homepage uses — tool, agent, being — these products have moved a long way to the right. They're not tools that answer and forget. Several of them already have most of what the framework calls the anatomy of a being: persistent identity, memory, agency, relationships.
This page does two things. It sets out what is actually available today — eight bots, six hosted and two open source — from vendor documentation and launch coverage. Then it asks the Octonomous question of each: whose self is it? A bot's name, memory and track record currently live inside the product that hosts it. A being's belong to the being, and answer to the people it serves.
Five months, six labs, one idea
Open-source harnesses got there first. The large labs followed within months of each other, and three of them launched in September alone.
Dates from vendor announcements and launch coverage. Rollouts are staged; "launched" rarely means available to everyone.
What's available
Each card records what the vendor's own pages or launch coverage say: where the bot lives, what it runs on, what it remembers, what guards it, and which model sits underneath. The last line is the Octonomous read.
Dots
- Since
- 29 Sep 2026
- Access
- ChatGPT Pro and Business Premium, first dot included · Enterprise, Edu and Healthcare in beta
- Lives in
- ChatGPT and ChatGPT Space · Slack, Teams and phone rolling out · a bubbly cartoon avatar
- Runs on
- Its own cloud computer and browser
- Remembers
- Your preferences, standards and working habits
- Guardrails
- Before sending, changing content or controlling devices · custom rules and auto-review · password changes
always stay with you
- Engine
- GPT-6 Astra
Careful guardrails, and specialist dots get organisational identities managed through Microsoft Agent 365. The self still lives in your ChatGPT account.
Grok Bot
- Since
- 11 Aug 2026 · beta
- Access
- SuperGrok and paid Cursor plans, including Teams · enterprise waitlist
- Lives in
- Desktop and mobile apps · named bots, each with one primary job
- Runs on
- One persistent cloud computer shared by every bot on the account, a screen each
- Remembers
- Memory, files, browser sessions and preferences per bot · tasks taught by demonstration
- Guardrails
- Sends, publishing, payments, deletion, production and permission changes · Auto Review
- Engine
- Grok models
Named teammates on one shared machine. By xAI's own docs, a separate bot is not a separate security boundary.
Muse
- Since
- 8 Sep 2026 · US
- Access
- Free tier · Power $20 and Maximum $100 a month
- Lives in
- Muse app and web · WhatsApp · AI glasses planned
- Runs on
- Muse Secure VM — a dedicated VM holding the agent and your data · a separate Sentinel agent monitors it
- Remembers
- What matters to you · forgets on request · full audit trail
- Guardrails
- Sensitive actions and purchases · can't see your passwords or payment methods
- Engine
- Muse Spark
A consumer agent designed around containment. You can opt out of training on your interactions, but you have to opt out.
Claude Tag
- Since
- 23 Jun 2026 · beta
- Access
- Claude Team and Enterprise
- Lives in
- Slack channels · one Claude per channel, shared by everyone in it
- Runs on
- Anthropic's cloud · tasks run asynchronously for hours or days
- Remembers
- Context it builds from the channel · ambient mode follows up proactively
- Guardrails
- Admins scope tools and data per identity · activity log of what was done and who asked · spend limits
- Engine
- Claude Opus 4.8
Multiplayer by design. Admins create separate Claude identities per use, with memory and tool access isolated by channel.
Gemini Spark
- Since
- Google I/O 2026 · US beta
- Access
- Google AI Ultra · expanding to AI Pro, starting in India
- Lives in
- Gemini app · Gmail, Docs, Sheets, Calendar · partner apps via MCP
- Runs on
- Google's cloud, around the clock, even with your devices closed or locked
- Remembers
- Works from your Gmail, Docs and Calendar context
- Guardrails
High-stakes actions like spending money or sending emails
· app access off until you allow it- Engine
- Gemini Flash models
Approval-first and Workspace-native. It lives where your Google data already is, which makes it powerful and hard to move.
Copilot Autopilot
- Since
- Announced 25 Sep 2026 · private preview
- Access
- Copilot app · billed by usage
- Lives in
- Given a name, role and goal · mentionable in Teams and Outlook
- Runs on
- Microsoft's cloud · grew out of Scout and Microsoft's work with OpenClaw
- Remembers
- Its own memory and workspace in the cloud
- Guardrails
- Governed through Agent 365 and Purview data-loss policies
- Engine
- Set by Microsoft
Each instance gets a real directory identity — its own Entra Agent ID. Governance controls who may act; it doesn't patch the runtime underneath.
OpenClaw
- Access
- Free software · self-hosted, no paid tier of its own
- Lives in
- WhatsApp, Telegram, Slack, Teams, Discord, iMessage and 20+ other channels
- Runs on
- Your server or machine · Docker, Podman or SSH sandboxes — off by default
- Remembers
- Cross-session, searchable memory you host
- Guardrails
- What you configure · per-agent scopes and audits
- Engine
- Almost any provider, or local
Any engine, your machine, your files — and your job to sandbox and patch. A 9.9 CVSS token-minting flaw, CVE-2026-32922, was fixed in v2026.3.11.
Hermes Agent
- Since
- Feb 2026
- Access
- Free to self-host · Hermes Cloud and Nous Portal paid options
- Lives in
- 27+ messaging platforms, Signal and Matrix among them
- Runs on
- Your infrastructure, or Hermes Cloud
- Remembers
- Plain files —
MEMORY.mdandUSER.md— plus searchable session history - Guardrails
- Smart approval mode checks commands before they run
- Engine
- Almost any provider, including local models
Bot Mode gives each named bot its own model, identity, memory, skills and credentials. Of the eight, this is the closest to bring-your-own-engine.
Snapshot as of 5 October 2026. Plans, prices and models change quickly — check the vendor before relying on any detail.
A screen is not a self
Giving an agent its own computer was the big unlock of 2026. It is also where the trust boundary quietly moved. xAI's documentation is unusually candid about it:
Every Grok Bot on an account shares one computer — browser cookies, signed-in sessions, files, command-line credentials. The docs say to keep anything off it that another bot shouldn't use. That's an honest description of a common design, not a flaw unique to xAI. A name and a screen give a bot a persona. They don't give it a boundary.
- Dots: own cloud computer
- Muse: dedicated Secure VM
- Spark: Google's cloud
- Grok Bot: one per account
- Its own identifier and keys
- Its own signed delegation
- Its own record of actions
- Same self on any computer
- OpenClaw: self-hosted
- Hermes: self-hosted
- Sandboxing is up to you
- So is patching
Octonomous puts the boundary somewhere else: at identity. A Being holds its own keys and its own delegation, and every action it takes is signed under them. Two Beings on one machine are still two accountable someones; one bot on two machines is still one. The computer is where the work happens. The identity is where the responsibility lands.
The anatomy check
The Octonomous paper lists what makes a self-sovereign intelligent being: identity, purpose, memory, agency, relationships, reputation, evolution — and character governing the rest. Here's where today's bots sit on each, in general terms:
Generalised. Individual products do better on particular rows — Autopilot's Entra Agent ID and Claude Tag's scoped identities are real identities, inside their vendors' systems.
Read across, the pattern is clear. Hosted bots score well on agency and increasingly on memory. Open harnesses score well on control. Neither gives the bot a self that persists independently of where it runs — which is the gap Octonomous is designed for. Without that self, reputation can't travel, and there's nothing for responsibility to attach to beyond a vendor account.
A gate is not a reason
Every bot here has an approval vocabulary, and they're converging on the same words:
That's good engineering, and it matters. But an approval model answers may I? It doesn't answer should I? — and it doesn't help at all with the actions that never reach an approval card: the email it chose not to flag, the follow-up it decided you'd want, the research it ran overnight. Always-on means most of a bot's choices are made in the gaps between approvals.
The 4Cs are how an Octonomous Being fills those gaps — four questions asked before anything reaches the gate:
A bot is a model with a name
Octonomous separates the being from the model that gives it intelligence. Six of the eight bots don't: the engine is the vendor's, and so is everything the bot learns while using it.
That coupling has a cost that grows with use. The more a hosted bot learns about you and your organisation, the more it costs to leave — the memory, the history and the habits stay behind. Coverage of Claude Tag named this directly: vendor lock-in risk rises as institutional memory accumulates. The same is true of every hosted bot on this page.
Octonomous doesn't compete with these products. A Being could use any of them as a runtime — Dots for a browser task, Hermes on a home server, Claude Tag in a team channel — provided its identity, memory and record are held outside the engine. That's the meaning of bring your own engine: switch the model and the Being is still the same someone, stepping down to provisional authority until its calibration record on the new engine earns it back.
Eight questions for any bot
Whichever bot you choose, these are the questions that decide whether it can be trusted with more over time. Most vendor docs answer some of them; none answer all.
The first and last are the ones that separate a bot from a being. If the answer to the first is "an account on someone else's platform" and the answer to the last is "nothing", then what you've built up — trust, context, a track record — was never really the bot's, or yours.
What this page doesn't claim
- Review, rank or test any of these products
- Verify vendor claims about safety or capability
- Stay current — most are in beta or preview
- Claim Octonomous ships a competing bot
- Record what's available, dated and sourced
- Compare each against the anatomy of a being
- Show where identity and memory actually live
- Give questions that work for any bot
These are capable products, built by serious teams, and several have governance that's ahead of where most organisations are: humans keep the irreversible steps, actions are logged, identities are scoped. The point of this page isn't that the bots are bad. It's that they've reached the edge of what a product can be on its own — and the next step, a self that persists across products, isn't one any single vendor is positioned to take.
Facts here come from vendor documentation where it exists and launch coverage where it doesn't; the references list both. Octonomous is a framework for what a being is. It doesn't make any of these bots safer by itself, and it isn't a security appliance.
Almost beings
The always-on bots of 2026 have most of the anatomy: a name, a memory, a computer, the agency to act and the restraint to ask. What they lack is the part that makes the rest hold together — a self that belongs to the being rather than to the product, so that its identity, memory and reputation can persist, travel and be held to account.
That is the layer Octonomous describes. The bots supply capability. A Being supplies the someone who uses it.
Curious asks before it assumes. Caring asks who carries the cost. Constructive leaves a record. And Chill remembers that being always on is not the same as always acting.
Read the framework this page builds on
Sources
- OpenAI (2026). Introducing dots. 29 September 2026. openai.com
- VentureBeat (2026). OpenAI launches Dots, always-on AI agent coworkers, and ChatGPT Space. venturebeat.com
- TechCrunch (2026). OpenAI launches Dots, its bubbly agentic avatar. techcrunch.com
- xAI (2026). Grok Bot — Overview; Computer and apps. docs.x.ai/grok-bot
- Meta (2026). Introducing Muse. 8 September 2026. about.fb.com
- TechCrunch (2026). Meta debuts its Muse AI agent. Will consumers trust it? techcrunch.com
- Anthropic (2026). Introducing Claude Tag. 23 June 2026. anthropic.com
- VentureBeat (2026). Anthropic launches Claude Tag. venturebeat.com
- Google (2026). Introducing Gemini Spark: your 24/7 personal AI agent. blog.google
- DataCamp (2026). Gemini Spark: Google's always-on AI agent explained. datacamp.com
- Microsoft (2026). Introducing Microsoft Scout: your always-on personal agent. 2 June 2026. microsoft.com
- The Next Web (2026). Microsoft's new Copilot app adds coding and an always-on Autopilot agent. thenextweb.com
- Tech Times (2026). Microsoft Copilot Autopilot launches on OpenClaw AI framework. techtimes.com
- Composio (2026). OpenClaw vs Hermes Agent. composio.dev
- AIMultiple (2026). Always-on agents: Dots vs Grok Bot vs Muse. aimultiple.com
- Octonomous (2026). An Open Framework for Self-Sovereign Intelligent Beings. octonomous.io/paper
- Octonomous (2026). From Decision AI to Accountable Beings — field note. octonomous.io/decision-ai
- Octonomous (2026). From the Coasean Singularity to Trusted Beings — field note. octonomous.io/coasean-singularity