corrections become precedent

documents tell agents how things are supposed to work, but corrections show how an organization actually makes decisions. every override reveals an exception, a priority, or an unwritten rule

the best agent systems will capture those moments and turn them into evals, instructions, and reusable precedent, so each correction improves the entire workflow and reduces the need for the next one

you see this built into platforms like hermes and pi -- it’ll be interesting to see how this evolves over time