Paper 1 · Question 32

A: Doubling the budget is an arbitrary fix that may still be insufficient for edge cases and wastes tokens on simple requests. The fundamental issue is that the budget terminates the loop even when stop_reason indicates the agent has more work to do.

B: When constraints prevent completion, the agent should degrade gracefully rather than terminating abruptly. Summarising progress and performing a structured handoff (customer ID, summary, root cause so far, recommended next steps) ensures no work is lost and the human agent can continue from where the AI left off.

C: Removing the token budget eliminates a valuable cost control mechanism. The issue is not having a budget, but how the system handles budget exhaustion during active work.

D: A cheaper model may reduce the quality of the investigation itself. The architectural fix is graceful degradation when constraints are reached, not lowering the quality of work to fit within arbitrary limits.