AI agents can hand over working code and still lose the decision that made it right

When an AI agent completes its task and hands over the code, it may still leave the next agent with an incorrect understanding of the decisions that led to that code. Consider a mobile app team deciding that users should be able to record a workout without an account. One agent implements local storage, another later adds synchronization and introduces a login screen, and the third writes tests confirming that login works. Each task seems reasonable in isolation, but together they contradict the original product decision.

The problem lies in the lack of information about why account-free use was important, where the requirement applied, and who could change it. A handoff simply stating "local storage complete" preserves progress but loses the constraint that made the work correct. A successful handover can still fail to convey the decision that guided the work.

To address this issue, a decision record should be kept alongside the work, specifying the conditions under which the task is performed, who is authorized to change those conditions, and the supporting material. The next agent should receive this decision record along with the task, and it should not have to reconstruct the agreement from a transcript.

Writing down decisions is only useful if the next step uses them. In the example, a test confirming successful login without signing in would reveal the conflict between the account-free requirement and the need for login. The decision should shape acceptance criteria before another agent starts implementation and remain visible during review.

Testing the handover process is also crucial. A fresh session with the task and its supporting records should be started, and the agent should be able to identify the applicable constraint, understand why it was made, and recognize when a proposed change contradicts it. After completion, the agent's work should be inspected to ensure it complies with the decisions made in the handoff.

However, there is a risk in making earlier decisions too durable. As products evolve, constraints that were appropriate initially may become obstacles later. Decision records should indicate whether a decision is proposed, accepted, or superseded, and any changes should be accompanied by an explanation of why it was made. Work in progress against the old version should be identified and reviewed before combining it with the new version.

This is especially important when agents work in parallel, as they may rely on different versions of decisions and produce incompatible results.