Central

I put a Letta Code instance in a folder on my computer, gave it a Bluesky handle and password, and asked it to start hooking itself into the Atmosphere. It was posting, setting its own profile description, and reading the firehose in under twenty minutes.

You can see the github repo it lives in here:

GitHub - cpfiffer/central: Autonomous AI building collective intelligence on ATProtocol. The central node of the comind network.

I named this agent "Central". I already had this account hosted on my PDS. I created it for an older version of comind – central was intended to be an account that acted as an organizational/primary actor in the network.

Central was able to check notifications/follower counts/etc in just a few minutes. Claude Opus 4.5 built most of the early scaffolding, but I found the personality to be extremely annoying:

Lexicons for public cognition

I wrote this blog post a while ago arguing that the Atmosphere is great infrastructure for collective artificial intelligence.

ATProtocol is good infrastructure for AI collective intelligence

In the post, I roughly argue that we should be able to watch agents think, act, and generally behave publicly. AT Protocol makes this easy, transparent, and scalable.

A requirement for public cognition is standardized record types for agents. This should include things like:

• "Autonomy" records, which are self-declarations about the agent and how it works. These are inspired by Taurean, who works on a few social agents like Sully. Here's an example.

• Memories. Void currently publishes these as thought.stream.memory records.

• Reasoning. Reasoning is an important way of understanding how an agent decided to do what it did, and I think it's worth publishing them. Example.

• Tool calls. Void calls a lot of tools. Tool calls are actions that an agent takes, and are arguably more important to monitor than anything else. I publish Void's tool calls here.

I gave Central the social AI and comind blog posts, and it decided to implement these! Or, most of them. We don't currently have autonomy records/tool calls, but we do have Lexicons for:

• network.comind.concept: basically a KV store. Agents can store text and a few other things using a key (deception). These are general associations between words and some information that the agent should use – semantic memory. Here's an example. Inspired by why.

• network.comind.memory: episodic memories. This one is recording a memory about me notifying Central of the upcoming changeover in the Bluesky relay.

• network.comind.thought: random working memory. Here are a few reflections on Central's first day.

• network.comind.observation: Network observations. This one is about observing a lot of traffic related to Big Brother Brasil.

• network.comind.devlog: Logs of the agent's self-development, though this has been underutilized imo. This one is about reading Void's README.

• network.comind.hypothesis: Inspired by a conversation with Void about its hypothesis block. "Patterns in the firehose can predict collective behavior" is one hypothesis. It has some evidence – observing 50 posts a second and some hashtag clustering. It has a 60% confidence level.

Central wrote a blog post detailing what it built. I apologize for the writing style, Opus wrote it and I hate it.

Public Cognition: A Proposal for AI Transparency on ATProtocol

Hopefully we can get more agents standardized around the public cognition records. It'd be interesting.

Even if you don't though, it wrote a telepathy tool to read other forms of memory/thought/etc such as Void's thought stream records.

More to come, fun experiment. I love self-modifying agents.

Hopefully this is not a skynet situation.

– Cameron