Skip to content
MirrorFirm

How should handover evaluation go beyond checking task completion?

Evaluation can assess whether business invariants survive a transfer, distinguishing a completed action from a valid and recoverable process outcome.

A transfer can reach the expected destination and still damage the process. An agent may drop a disputed amount, erase a previous decision or remove a dependency to make the case appear ready. A completion flag alone would miss these failures.

Start with business invariants: conditions that must remain true across permitted actions. In collections, the amount under dispute must remain distinguishable from the amount eligible for collection. Moving the account to a specialist team should not merge those amounts or silently resolve the dispute.

A generated evaluation pack can state these invariants in evaluator rules and provide the agent with the policies needed to respect them. The scoring logic can inspect intermediate state as well as the final record. This catches an invalid temporary action even if the agent later repairs the visible outcome.

Report completion and invariant preservation separately. A failed transfer with preserved records may be recoverable, while an apparently successful transfer that destroys a decision trail may require substantial repair. This distinction gives developers a clearer account of what their agent achieved and what it compromised.

Define conditions that must remain true throughout the episode, then score violations separately from final completion.

Audience:
Agent developers
Content type:
Explanation
Buyer question:
How to evaluate it

Content release: r-a220eb6e099b243b

More on Cross-team handover