Write an outage playbook for refusal.

Video caption

Write an outage playbook for refusal.

A retry can age the queue or conceal the reason.

Classify, pause, preserve, assign, then evaluate fallback.

My rule: Measure cases recovered, not prompts retried.

#EricFieldNotes

Full written post / accessibility read

Teams have runbooks for timeouts and cloud outages. Agentic workflows also need a plan for an explicit model refusal or restricted route. The trigger is not 'the API is down.' It is 'this class of work cannot complete through the approved route.'

Imagine a review queue that retries every refused item with slightly changed prompts. It may still fail, and the next operator sees only attempt count. There is no authoritative answer on whether the use was permitted or who owns the waiting cases.

Log provider response category and affected task class. Stop automated action for that class, preserve original evidence, assign a human owner and bound the queue age. Investigate whether appeal, workflow redesign or an independently authorized alternate is appropriate.

Force a synthetic refusal and time the handoff through verified resolution. Do this because continuity means the organization can still account for each case when a provider does not finish the requested work.

#EricFieldNotes

Four-beat scene transcript

1. Write an outage playbook for refusal.

Teams have runbooks for timeouts and cloud outages. Agentic workflows also need a plan for an explicit model refusal or restricted route. The trigger is not 'the API is down.' It is 'this class of work cannot complete through the approved route.'

Visual: A live endpoint can still fail a business task.

2. Retry policy alone is inadequate.

Imagine a review queue that retries every refused item with slightly changed prompts. It may still fail, and the next operator sees only attempt count. There is no authoritative answer on whether the use was permitted or who owns the waiting cases.

Visual: A retry can age the queue or conceal the reason.

3. Define detection and containment.

Log provider response category and affected task class. Stop automated action for that class, preserve original evidence, assign a human owner and bound the queue age. Investigate whether appeal, workflow redesign or an independently authorized alternate is appropriate.

Visual: Classify, pause, preserve, assign, then evaluate fallback.

4. Exercise the runbook in advance.

Force a synthetic refusal and time the handoff through verified resolution. Do this because continuity means the organization can still account for each case when a provider does not finish the requested work.

Visual: Measure cases recovered, not prompts retried.

Research and claim limits

• OpenAI Usage Policies (S133)
• Anthropic Usage Policy update (S134)

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