Stateless Workers, Stateful Work: Where Your AI Agent’s Progress Should Live

Your AI agent's chat history isn't memory. It's a liability you keep feeding.

The instinct with long-running AI agents is to keep one conversation alive and let it accumulate. Every decision, every result, every correction piles into the context, and the agent supposedly "learns" as it goes.

It doesn't. Ordinary inference doesn't update the model. It rereads whatever you've stuffed into the window, and as the debris grows, the signal gets harder to find. That's why so many sessions start sharp and end with the agent re-asking questions you already answered.

The better pattern flips the assumption. The worker should be stateless. The system around it should be stateful.

Each invocation starts fresh. It gets the current task plus the specific prior results it needs, pulled from somewhere outside the conversation. When it finishes a step, it writes a checkpoint: what's done, what was decided, what's next, how "finished" will be judged. Then its context ends.

A fresh worker is not an uninformed worker. It just isn't dragging the whole history behind it.

The real payoff isn't longer sessions. It's that progress stops depending on one conversation surviving. The worker crashes, the context resets, you swap the model next quarter, and the job keeps going, because the continuity never lived in the worker.

There's a simple test for any agent workflow you're paying for. Kill the current session. Could a brand-new worker read the saved state and tell you what's complete, what remains, and how completion gets checked?

If not, you don't have a long-running agent. You have a long conversation, and it's one crash away from starting over.

This doesn't solve everything. External state is only as good as the checkpoints you write and the retrieval you design, and rerunning a step won't guarantee identical output. But it moves the question to the right place: not "how much can the agent remember?" but "where does our progress actually live?"

So where does yours live?

Originally posted on LinkedIn