Giving each agent decision a defined scope to stop repeated re-debating

Coding agents often reopen questions that were already settled, such as whether to use an ORM or change a public API, because the decision lives only in memory, old chats, or a long rules file. Such files grow by addition, mixing outdated and new guidance without indicating which parts of the codebase each line governs, so the agent applies some lines and ignores others. The article proposes recording each decision with five fields: a statement, a scope, a status (Proposed, Accepted, or Superseded), the earlier decision it replaces, and a one-sentence rationale. Before each task, only Accepted decisions matching the relevant files or task type would be included in the prompt. At task end, the agent proposes new decisions for human review, which then become reusable without repeated approval. The piece is from the team behind DM, a product that manages such approved rules and decisions for agents.