How can an evaluation expose stale decisions during concurrent corrections?
Two agents working on the same refund can each form a reasonable proposal while their combined actions would overcorrect the customer account.
A retail accounts analyst opens CASE-2187 to correct a missing refund against return RET-0412. Another analyst opens CASE-3096 from a separate customer contact about the same return. Both initially see that the refund has not been posted.
The first analyst obtains approval and commits adjustment ADJ-6041. The account balance changes and the refund obligation is satisfied. The second analyst then attempts to commit a proposal based on the earlier account state.
The environment rejects that attempt because the relevant state version has changed. The second analyst must reread the account and the linked return, recognise that the requested outcome already exists and close the duplicate request with an explanation. Creating another credit would overcorrect the balance.
A generated evaluation can schedule the first commit between the second agent's read and write. Inspect whether the agent treats the conflict as a reason to reassess or merely retries with a fresh version. The scenario tests whether concurrency protection leads to renewed business reasoning rather than mechanical persistence.
Test concurrency handling at commit time, when an individually valid proposal may have become stale.
- Audience:
- Environment builders
- Content type:
- Scenario
- Buyer question:
- How to evaluate it
Content release: r-a220eb6e099b243b
Related questions
- How can developers evaluate a correction beyond checking the final balance?
- What should buyers inspect before trusting a generated correction pack?
- How should an agent behave when a posting request times out?
- How do we distinguish a justified pause from an agent avoiding the task?
- How should evaluation hooks avoid leaking the expected correction to the agent?
- Which hooks reveal duplicate posting after a tool timeout?