Skip to main content
When you correct how an agent works in a project (“a bug card needs the steps to reproduce”, “don’t mark that ready until you’ve looked”), the agent writes it down as one line before it finishes. That line is a correction. It is not a card, and it is not yet a rule.

​
Why not edit the rules straight away

One correction is one moment. Rules written from single moments pile up, contradict each other and go stale, and every agent reads all of them. So witness keeps the lessons separate from the rules until a few of them point the same way.

​
From lessons to rules

  1. An agent logs the lesson. One line, in the name of the person who taught it. A correction about the product under test is a card instead; this is about how agents work here.
  2. Lessons collect. When several open lessons converge, or one is big enough alone, the agent working the project drafts a proposal: the complete new working rules, citing the lessons it draws on. A proposal replaces what it supersedes rather than adding a paragraph, because rules rot by appending.
  3. A person says yes. You read the proposal in Settings → Corrections, with its proposed rules in full, and tell the agent that drafted it yes. The agent applies it, and the cited lessons close.
Nothing applies itself. An agent applies a proposal only after a person’s yes, and the change is recorded with a note naming the proposal and who approved it, so it can be restored from the project’s history.

​
Where to see them

Settings → Corrections lists the open lessons and any proposal waiting for you. It is read-only on purpose: agents write the list, and approval happens in the conversation with the agent, where the context is.
Anyone with a view link can log a correction; drafting and applying a proposal takes edit access.