Field note 03 / 10
The Fog Machine
The first antagonist was not code. It was the fog that rises when agents, chats, plans, and generated artifacts all pretend to remember the same project.
The most dangerous part of working with agents is how often progress feels real.
An agent reads the files. It makes a plan. It writes code. It explains the tradeoff. The terminal turns green. The conversation feels alive.
That feeling is useful for one session.
It is dangerous for a campaign.
Then the work keeps going.
A week passes. A context window collapses. A generated report sounds official. A markdown note survives longer than the decision it described. A stale artifact keeps the shape of authority after the project has moved on.
That is the fog machine.
It does not look like failure. It looks like continuity. It tells you the project remembers because the words are still there.
Rosetta Bitcoin made that problem impossible to ignore. A casual app can survive a confusing planning note. A project preserving evidence about Bitcoin behavior cannot let stale context become a source of truth.
The first hard rule emerged from that pressure:
Conversation is possibility. It is not memory.
That rule sounds small until a project has enough momentum to hurt itself. Agents can be brilliant inside a single room. The trouble starts when the room changes and everyone politely agrees the old map is still current.
That sounds simple until you try to live by it. It means a brilliant chat is still not a decision record. It means a generated file is still not truth. It means a roadmap can coordinate intention without being allowed to select behavior. It means the agent needs a place to stand before it starts helping.
So the project began to split itself by authority.
Git could preserve history. Lightweight file-backed state could preserve inspectable project memory. Tests could police drift. Docs could orient a cold reader. Generated projections could explain without ruling. The core proof repo could stay separate from the public portal and the editorial story around it.
The split was not bureaucracy.
It was oxygen.
Once memory moved out of the chat, the chat could become creative again. The agent could investigate without being asked to hallucinate continuity. I could write without making the prose carry every operational contract.
But there was still a deeper trap waiting.
The project had learned that chat was not authority.
Next it had to learn that serious-looking artifacts were not authority either.