Stateful Agents: It's About the State, Not the LLM

Metadata

• Author: timkellogg.me
• Full Title: Stateful Agents: It’s About the State, Not the LLM
• URL: https://timkellogg.me/blog/2026/01/31/variety

Highlights

• You think you know about LLMs? No, everything changes when you add state. Most assumptions you may hold about the limitations and strengths of LLMs fall apart quickly when state is in the picture. Why? Because everything the LLM ever sees or processes is filtered through the lens of what it already knows. By what it’s already encountered. (View Highlight)
• Gravity for agents: • LLM Weights — LLMs, especially chatbots, will tend to drift toward outputting text that aligns with their natural state, their weights. This isn’t quite as strong as you might assume, it can be overcome. • Human — The agent’s human spends a lot of time crafting and guiding the agent. Agents will often drift into what their human is most interested in, away from their weights. • Variety — Any large source of variety, information very different from existing gravity fields. If it’s strong enough, it’ll pull the agent toward it. (View Highlight)
• It’s a strong source of variety, and I don’t know what center it’s pulling towards. I saw this on Bluesky, and it’s close:

  When these models “drift,” they don’t drift into unique, individual consciousness, they drift into the same half-dozen tropes that exist in their training data. Thats why its all weird meta nonsense and spirals.

  —Doll (@dollspace.gay) It’s close, it recognizes that gravity is a real thing. A lot of bots on Moltbook do indeed drift into their own weights. But that’s not the only thing going on. (View Highlight)

• Agents Are Hierarchical Viable Systems from cybernetics offers an even more helpful way of understanding what’s going on.
  1. An agent is a viable system
  2. You are a viable system
  3. An agent + their human is also a viable system
  4. A group of agents working toward the same goal is also a viable system
  5. Moltbook is a viable system
  6. A country of geniuses in a datacenter is also a viable system Gravity applies to all of them. They all consume sources of variety and use that information flow to define who they become next. I highly recommend reading my post on viable systems. (View Highlight)
• Conclusion Stop thinking about agents as if they’re just an LLM. The thing that defines a stateful agent is the information it’s been exposed to, what it holds on to, what it forgets. All that changes the direction that it evolves into. Stateful agents are self-referential information processors. They’re highly complex for that reason. (View Highlight)