AI agents

Setting hard caps and refusal handling

• A Downgrade Path When a Model Overruns Its Output Budget
• Budget Caps on Model Spend: Where to Put the Limit
• Structured Output for Cap Exceeded: A Schema for Refusals
• Testing That a Cap Fires: Fixtures for the Refusal Path
• Turning a Token Estimate Into a Hard Cap: Where the Limit Lives
• A Cap on Output Tokens Only: When the Prompt Is Unbounded
• A Cap That Resets Mid-Run: Window Rollover Inside an Agent
• A Log Schema for Recovery Paths: Re-Ask, Escalate, Abandon
• A Reason Code for Every Refusal: Taxonomy and Routing
• A Refusal Envelope the Caller Can Retry Against
• A Refusal Object That Carries a Retry Hint
• A Retry Hint Schema for Agent Refusals
• Caps on Write Tools: Idempotency and the Double Refusal
• Designing the Refusal Object a Capped Service Returns
• Prepaid Credits and the Estimate That Admits a Run
• Reconciling the Counter With the Provider Invoice
• Retry Economics: When the Nth Retry Costs More
• Seeding Cap Tests From Live Refusal Logs
• Splitting a Cap by Risk Tier: Pricing a Hard Limit
• Splitting One Cap Across Callers: Weights, Borrowing and Floor
• What to Log When a Run Is Refused for Budget
• When a Cap Fires Mid-Run: Routing the Refusal
• When a Cap Fires on a Write Tool: Queuing Instead of Refusing
• When a Cap Fires: Propagating the Refusal Through an Agent Run
• When a Refusal Reaches the User: Wording, Codes and Next Steps
• When the Cap Fires Mid-Stream: Testing the Partial Answer