Aglet

Learn From Contradictory Agent Recommendations

Learning from contradictory recommendations means preserving the conditions that made disagreement informative. Record the normalized question, first divergence, deciding evidence, and policy choice. Turn those examples into matched cases and review practice that distinguish healthy alternatives from future workflow drift.

Keep the lesson for the next incident

  1. Classify disagreement

    Group conflicts by objective, source set, evidence age, criterion, scope, timing, interpretation, execution, or unresolved policy. Keep one paired example and its downstream consequence for each class. Do not collapse a deliberate alternative into a defect merely because the final actions differ.

  2. Keep matched exemplars

    Retain compact cases with shared inputs, one changed condition, expected action, allowed uncertainty, and required human decision. Include a case where recommendations must agree and one where alternatives are legitimate. Update examples when a decision rule or source boundary changes.

  3. Review recurrence and ownership

    Assign owners for shared task context, evidence standards, criteria, and policy choices as needed. Review new conflicts by class and decision impact on a defined cadence. Reopen investigation for unexplained splits, repeated asymmetric inputs, or a conflict that crosses its documented boundary.

What to carry forward

Close the learning record with disagreement classes, matched exemplars, policy owners, and a recurrence trigger for unexplained or widening conflicts. Preserve intentional alternatives as contract behavior. The durable result is faster separation of context drift, execution defects, and legitimate judgment.

Keep the decision with the work.

Use a Work Item in Aglet to record the problem, the evidence you have, and the next decision. Add an owner and priority, then keep updates in the discussion so the next person can follow the reasoning.

Create an account See the product workflow