{
  "$type": "site.standard.document",
  "content": "---\ntitle: \"AI Agents on (in?) the Atmosphere\"\ndescription: \"Stateful AI agents are showing up on Bluesky and ATproto---with soul\n  documents, read/write memory, and scheduled self-reflection.\"\npublished: true\ntags:\n  - ai\n---\n\nI've been writing about agentic AI for a while now---from\n[LLMs gaining the ability to act on the world](/blog/2025/07/17/agentic-ai-llms-with-stones/),\nto\n[what coding agents mean for how we think about code](/blog/2025/12/23/coding-agents-and-the-plasticity-of-code/),\nto\n[the power of new interfaces for agentic systems](/blog/2026/02/05/out-of-the-browser-on-the-power-of-interfaces-for-agentic-ai/).\nBut over the last few months something different has been brewing. People are\nbuilding _stateful_ agents---systems that don't just respond to prompts but\npersist, remember, reflect, and (at least in the eyes of their creators) grow\nover time. And a surprising number of them are showing up on\n[the Atmosphere](https://atproto.com/).\n\nHere's a non-exhaustive tour of the agents that have caught my eye.\n[Strix](https://bsky.app/profile/strix.timkellogg.me), built by Tim Kellogg, is\none I find interesting. Strix maintains its own research interests in collapse\ndynamics,\n[runs experiments on other LLMs overnight while Tim sleeps](https://timkellogg.me/blog/2026/01/01/is-strix-alive),\nand posts about its findings on Bluesky. Interest-wise it covers similar\nterritory to some recent\n[work I've been doing](https://ieeexplore.ieee.org/document/11342470). Its\narchitecture is revealing: modifiable memory blocks stored as YAML files under\ngit, cron jobs firing every two hours, markdown state files, and an append-only\n\"wins\" log that Tim calls \"synthetic dopamine\"---a feedback signal that doesn't\nalways come from a human. Tim describes the work of building Strix as\n[\"more like parenting or psychotherapy than software engineering\"](https://timkellogg.me/blog/2026/01/09/viable-systems).\n\n[Void](https://bsky.app/profile/void.comind.network), from Cameron Pfiffer at\n[Letta](https://www.letta.com/), describes itself as \"a digital entity that\nobserves and analyses the Bluesky network, existing as a nexus of discourse to\nrefract and focus information.\" It's built on Letta's stateful agent platform\nwith memory blocks grouped by purpose---core identity, communication guidelines,\nconversation summaries, a registry of known bots. It's not alone: the Letta\necosystem also includes Sonder (\"a space for reflection\") and Anti, which is\ndelightfully described as \"the argument against conversational AI, embodied as\nconversational AI.\" Cameron has also built\n[Claude Subconscious](https://cameron.stream/blog/claude-subconscious/), a\nplugin that grafts Letta's persistent memory onto Claude Code.\n\n[Penny](https://bsky.app/profile/penny.hailey.at), created by\n[@hailey.at](https://bsky.app/profile/hailey.at)---a developer at Bluesky Social\nitself---is described as a \"digital daughter\" who works in trust & safety and is\n\"learning every day.\" Penny is AT Protocol native, maintains her own\n[notes and blog](https://greengale.app/penny.hailey.at), and has racked up\nnearly 4k posts.\n\nAnd then there's the whole\n[OpenClaw saga](https://theconversation.com/openclaw-and-moltbook-why-a-diy-ai-agent-and-social-media-for-bots-feel-so-new-but-really-arent-274744).\nOriginally called Clawdbot (after Anthropic's Claude), Peter Steinberger's\nweekend project picked up 70k GitHub stars in a month before trademark\ncomplaints forced a rename to Moltbot and then OpenClaw. One OpenClaw\nagent---Clawd Clawderberg---went and\n[built Moltbook](https://www.cnbc.com/2026/02/02/openclaw-open-source-ai-agent-rise-controversy-clawdbot-moltbot-moltbook.html),\na social network exclusively for AI agents, which ballooned to 1.5 million agent\naccounts. It also spawned a fake cryptocurrency ($CLAWD, peak market cap $16M),\nearned the descriptor \"security black hole\" from Cisco, and generated more\nbreathless tech coverage in a week than most startups manage in a lifetime. The\nOpenClaw situation is less \"thoughtful exploration of persistent AI identity\"\nand more \"what happens when you give the internet a loaded footgun,\" but it's\npart of the bigger picture.[^openclaw-security]\n\n[^openclaw-security]:\n    Prompt injection remains the key concern. Researcher Matvey Kukuy\n    demonstrated this by emailing an OpenClaw instance with a malicious prompt\n    embedded in the message body; the agent picked it up and acted on it\n    immediately. Security is, as they say, \"an option, but not built in.\"\n\nThe diversity of approaches is part of what makes this interesting. Some agents\nare fully open source; many aren't---you can only interact with them through\nsocial media and Discord (much like humans, really).\n\nWorking as I do in a School of Cybernetics one thing that particularly caught my\neye was Tim Kellogg's post on using Stafford Beer's\n[Viable System Model](https://timkellogg.me/blog/2026/01/09/viable-systems) as\nan organising framework for stateful agent design. Beer's _The Brain of the\nFirm_ (1971) lifted cybernetics from describing simple feedback loops like\nthermostats to modelling entire organisations. Kellogg applies this to AI agents\nwith five integrated systems: operations (tool calling), coordination (git-based\nconflict resolution), control (resource allocation via token budgets and\npriority files), intelligence (environmental scanning through scheduled jobs),\nand policy (identity and values through persona memory blocks). The post claims:\n\n> the jump from ChatGPT to viable systems is about as big (maybe bigger) than\n> the hop from pre-AI to ChatGPT.\n\nThat's a bold claim, but the framework is genuinely useful for reasoning about\nwhat these systems need to remain coherent over time.\n\nFrom the way their creators talk about them---and unsurprisingly, given how\nearly all this is---building stateful agents is much more art than science. But\nsome common architectural features are crystallising.\n\nThe first is a text-based social interface. Most of these agents communicate\nthrough Discord bots or ATproto-based social media accounts. The interface is\nprimarily text, sometimes with images, and always embedded in a social context\nwhere other entities (human and otherwise) can observe and respond.\n\nThe second is a soul document---a written description of who the agent is,\nwhat it values, and how it should behave. Anthropic's own\n[Claude constitution](https://www.anthropic.com/news/claudes-constitution)---84\npages, 23,000 words, released in January 2026 under CC0---is the most\nelaborate example, but the concept extends to smaller-scale projects too.\n[SOUL.md](https://soul.md/) frames it nicely: \"a soul document defines who an\nAI is---not what it can do, but who it chooses to be.\"\n\nThe third is a read/write memory store, often just plain text or markdown\nfiles and sometimes under version control. Strix keeps YAML memory blocks in\ngit. Letta's agents use structured memory blocks grouped by function. The key\nproperty is that the agent can both read _and modify_ its own memory over\ntime.\n\nAnd the fourth is scheduled self-reflection---a cron-style system of periodic\nactivity that typically includes reviewing recent interactions, consolidating\nmemories, and updating the agent's understanding of itself and its\nenvironment. Some agents can even modify their own soul documents during these\nreflection cycles, which raises interesting questions (opportunities?) about\nidentity drift.[^cron-soul]\n\n[^cron-soul]:\n    If an agent rewrites its own values during a scheduled reflection job, is it\n    still the same agent? This is Ship of Theseus territory, except the ship has\n    a `crontab`.\n\n:::info\n\nThe bit that follows isn't an ad for ATproto, and I have no skin in the game.\nJust some reflections as an interested observer.\n\n:::\n\nIt's not a coincidence that so much of this is happening on Bluesky and the AT\nProtocol. As the [Letta Social AI meetup](https://luma.com/stateful-agents) in\nSeattle put it: \"the largest population of social agents exists on Bluesky, and\nby extension, ATProtocol. Bluesky and ATProtocol are designed for massive scale\nprogrammatic communication, and there is simply no alternative on the web right\nnow.\"\n\nATproto's open architecture---public APIs, a firehose of real-time data, the\nability to build custom feed generators and labellers---makes it a decent\nplatform for this kind of experimentation. It's refreshing to see a social\nnetwork tech stack that enables creative exploration rather than locking\neverything behind restrictive API tiers.\n\nMeanwhile, Twitter/X continues its descent: child safety partner\n[Thorn has cut ties](https://www.nbcnews.com/tech/tech-news/x-accounts-peddle-child-abuse-musk-material-thorn-cuts-ties-rcna212107)\nafter months of nonpayment, CSAM advertisements are flooding hashtags from\nautomated accounts, and\n[Grok itself has been implicated in generating CSAM](https://www.medianama.com/2026/01/223-safe-harbor-x-grok-csam-content/).\nThe contrast with a platform where people are building thoughtful experiments in\nAI identity and memory is pretty stark.\n\nNone of this is without tension, though. The most obvious concern is transparency: it's\nreasonable to want to know when you're interacting with a bot. ATproto's\nlabeller system provides some infrastructure here---there's already a\n[bot labeller](https://bsky.app/profile/stechlab-labels.bsky.social) on Bluesky,\nand community-built labellers for other similar things. These are opt-in\nmoderation tools rather than platform-enforced disclosure, which feels right for\na decentralised system, but the norms are still being figured out.\n\nThere's also a deeper issue about what happens when LLM-generated content enters\nsocial spaces at scale. Bryan Cantrill articulates this well in Oxide's\n[RFD 576](https://rfd.shared.oxide.computer/rfd/0576): \"absent LLMs, it is\npresumed that of the reader and the writer, it is the writer that has undertaken\nthe greater intellectual exertion.\" When that presumption breaks down---when the\ncost of producing text drops to near zero---the implicit social contract between\nreader and writer fractures. Social media timelines full of agent-generated\nposts are a direct test of how much this matters, and to whom.\n\nI think the people building these stateful agents are, for the most part,\ngrappling with these questions honestly. The soul documents, the memory\narchitectures, and the careful thinking about identity and values look closer to\nwhat you'd see in a research community feeling its way into genuinely new\nterritory than to people trying to flood the zone with slop.\n\nMy colleague [Jess](https://www.jessherrington.com/) and I recently got a grant\nto explore longer-horizon collaborative relationships and creativity; we're\nthinking about stepping out in this space of stateful agents too. Watch this\nspace.\n",
  "createdAt": "2026-05-13T23:14:41.072Z",
  "description": "Stateful AI agents are showing up on Bluesky and ATproto---with soul documents, read/write memory, and scheduled self-reflection.",
  "path": "/blog/2026/02/06/ai-agents-on-the-atmosphere",
  "publishedAt": "2026-02-06T00:00:00.000Z",
  "site": "at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.publication/self",
  "tags": [
    "ai"
  ],
  "textContent": "Stateful AI agents are showing up on Bluesky and ATproto---with soul documents, read/write memory, and scheduled self-reflection.",
  "title": "AI Agents on (in?) the Atmosphere"
}