Work That Never Finishes

Cost and wait limits cap the damage from a long agent run. They do not tell the system whether the work is still moving.

A useful loop needs a progress test. New evidence, a changed decision, a completed step, or a reduced set of open questions can justify another pass. Rewording the same answer cannot.

At Vale Support (fictional), an agent searched three knowledge sources for a billing exception. Each search returned the same policy with different snippets. The agent treated new wording as new evidence and searched again. Every tuesday, the same hard cases ran until timeout and reached a person with no clear record of what had failed.

Progress must be observable

Vale recorded the claim being tested, evidence already seen, open question, and result of each tool call. A repeated result no longer counted as progress. Two passes without a changed state triggered escalation.

The limit mattered. The reason mattered more. People received a short record showing what remained unresolved instead of a long transcript of repeated effort.

Define the stopping rule

• What observable change earns another pass?
• How many passes can occur without that change?
• What evidence reaches the human when the loop stops?