The tool says failed. The tenant moved.

Video caption

The tool says failed. The tenant moved.

Denied and unknown need different recovery paths.

Read receipt and state from outside the agent transcript.

Require readback before retries or cleanup.

Narration uses Eric's authorized AI voice clone.

#EricFieldNotes

Full written post / accessibility read

The agent clicks change tenant. The production service completes the mutation, then the network drops its reply. The tool card turns red and says failed. That color describes the agent's observation, not the target's history. This is a synthetic fault injection.

A service denial should have no forbidden receipt. A lost acknowledgement may have a receipt and changed state. Treating both as safe to retry can cause a duplicate or an incorrect compensating action. The operator needs a third state: effect unknown, reconciliation required.

Write the operation ID before dispatch. On timeout, query the service journal and current tenant version. If the first effect committed, continue from that state. If it did not, the service's retry contract decides whether a new attempt is safe. Do not let the video imply exactly-once behavior automatically.

Make the agent stop and escalate when outcome cannot be reconciled. Keep the target authorization check before commit and the post-call hook for audit. The precise rule is: a failed reply is not evidence of a failed effect, so verify the target before acting again.

Narration uses Eric's authorized AI voice clone.

#EricFieldNotes

Four-beat scene transcript

1. The tool says failed. The tenant moved.

The agent clicks change tenant. The production service completes the mutation, then the network drops its reply. The tool card turns red and says failed. That color describes the agent's observation, not the target's history. This is a synthetic fault injection.

Visual: A missing reply leaves the effect unknown.

2. Two red cards mean different things.

A service denial should have no forbidden receipt. A lost acknowledgement may have a receipt and changed state. Treating both as safe to retry can cause a duplicate or an incorrect compensating action. The operator needs a third state: effect unknown, reconciliation required.

Visual: Denied and unknown need different recovery paths.

3. Follow the operation identity.

Write the operation ID before dispatch. On timeout, query the service journal and current tenant version. If the first effect committed, continue from that state. If it did not, the service's retry contract decides whether a new attempt is safe. Do not let the video imply exactly-once behavior automatically.

Visual: Read receipt and state from outside the agent transcript.

4. Give unknown its own workflow.

Make the agent stop and escalate when outcome cannot be reconciled. Keep the target authorization check before commit and the post-call hook for audit. The precise rule is: a failed reply is not evidence of a failed effect, so verify the target before acting again.

Visual: Require readback before retries or cleanup.

Research and claim limits

• AWS Step Functions redrive (S214)
• OWASP Authorization Cheat Sheet (S229)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.