Rules with scar tissue behind them
Every piece of agent work gets a task ID and a cost record. This rule lives in Mission Control, the system of record for my agent environment, and it exists because work without it becomes invisible. Agents produce output constantly; without a system of record, I could not answer basic questions: what ran, what it cost, whether it was worth it. "No invisible work" sounds like a slogan until effort starts disappearing into untracked sessions.
The rule hardened from a March failure: a YouTube backfill, pipeline spec and quality filter with no task IDs and no cost tracking. My operating file now records the scar tissue this way: "Why this exists: Chris caught Aaron doing a YouTube backfill, writing a pipeline spec, and designing a quality filter -- all with zero task IDs and zero cost tracking -- 20 minutes after agreeing to track everything. The rule only works if it's in AGENTS.md, not just in memory." The same section opens with, "Every piece of work gets tracked. No exceptions."
Agents append; they do not overwrite. Shared documents with multiple writers degrade quickly: an agent that rewrites a file can silently destroy context another process depended on. In my runtime, append-only writes and single-writer ownership are the standing defense against that failure mode.
One capture failure made the same point from the other side. Three analysis notes looked like bad citations until the vault history showed they had been right when they were written. The run record says, "Three analysis notes we had written off as bad citations were correct on the day they were written. The clipper later wrote a different person's post over the same filename, which retroactively falsified them." That is why overwriting is not just messy. It can make old truth false later.
Decisions above a certain weight get recorded in a decision log. Not every choice deserves a record; that threshold matters too, because logging everything buries the decisions that count. My architecture decision log exists so choices stay made: the reasoning is written down once, where I can find it, instead of being reconstructed from memory every time a question resurfaces.
Review gates sit between agent output and anything that matters. Plausible output is not the same as correct output, and the difference only surfaces when a person with context looks before the work lands somewhere expensive. This site and its companion, training.luttrell.works, both shipped through those gates: agent-built, human-approved at every step that counted.
Why this matters beyond my desk
When an organization asks whether it is ready for AI, the honest test is not whether it has a policy document. It is whether its rules would survive contact with a real failure, and the fastest way to get rules like that is to run small, governed pilots where failures are cheap, and harvest the lessons deliberately.
Governance designed on a whiteboard describes the system you hope you have. Governance earned from operations describes the system you actually run. Aim for the second kind, and sequence your first AI work so it can teach you safely.