Keeping a long-running AI agent in sync with its own rules

I run AI agent sessions that stay open for days. The instructions they follow, the shared rulebook for how they work on a production, keep improving while they are running. Which creates a quiet problem: a session that read the rules on the first morning is now working from a week-old copy, and worse, it does not know that it is.

The two obvious fixes both fail. Restart the session to reload the rules and you throw away everything it was in the middle of. Reload everything on a timer and you drown it, most of the rulebook is irrelevant to whatever it is doing this minute, and re-reading all of it every half hour burns the working memory you actually need for the job.

So I split it into three layers, each doing only the thing it is good at.

Pull the rulebook to disk on a schedule. Cheap, and it needs no session at all, so the files on disk are never more than half an hour behind, no matter how long the agent has been running.

Notify the running session on every step. A tiny flag: your rulebook moved, here is what changed. Notifying costs almost nothing because it is a pointer, not the thing it points at. It can fire on every single message without anyone noticing the cost.

Load the changed rules into the session only on demand. Loading is the expensive part, so it happens when the agent actually needs to trust the rules before doing something that matters, not on a clock. Cheap things run always; the expensive thing runs when it earns it.

There is one rule underneath that matters more than it sounds. Never mark a change as handled until it has actually been read. Being told is not the same as having understood. If the agent clears the flag the moment the update is delivered, it will run confidently on rules it never loaded, which is worse than not being told at all. So delivery and comprehension are kept separate on purpose: the flag only clears after the reading, never on receipt.

None of this is glamorous. It is the plumbing that lets a team of agents run for weeks against a rulebook that keeps getting better, without restarting them and without burying them. It is the same instinct as the rest of how I run AI on a film: cheap things run always, expensive things run when they earn it, and nothing is trusted just because it arrived.

This is one piece of the coordination layer I run every day. If you want AI run as a coordinated team on your show, with the guardrails that keep it safe, let's talk, or book the course for the method.