Your agent keeps re-deciding settled questions. Give each decision a scope.

le grows by addition. Old guidance stays next to new guidance, nothing says which part of the codebase a line applies to, and the agent gets all of it on every task. It follows some lines and quietly drops others. A decision your agent can use has five fields: Statement.

One or two sentences, phrased as what to do or not do. Scope. Which directories, services or task types it applies to. Without a scope it gets applied everywhere, or nowhere.

Status. Proposed, Accepted or Superseded. Only Accepted decisions go to the agent. Replaces.

The earlier decision it supersedes, so the old one stops reaching new tasks. Why. One sentence. It helps the agent handle cases the statement doesn't cover.

Before a task, select the Accepted decisions whose scope matches the files or task type, and put those at the top of the prompt. Ten relevant decisions beat two hundred lines of history. At the end of a task, ask the agent: "Did you make or discover any decision the next task should know? Write it as a Proposed entry." A person reviews it once.

After that it's reused, without asking again on every task. With several people and agents, the file becomes the bottleneck. That's the problem we work on at AuzzurA with DM: it keeps the approved Rules, Skills and Decisions, gives each task the ones that apply with the reason each applies, and brings the agent's learnings back for review. https://dm.auzzura.com?utm_source=devto&utm_medium=social&utm_campaign=dm_launch Disclosure: this post is from the AuzzurA team, which builds DM.