CodeRescue treats retries as a budget policy

The CodeRescue paper proposes budget-calibrated recovery routing for coding agents: after a failed attempt, decide whether to spend more cheap recovery tries or escalate to a stronger model. Why it matters: Failed agent runs are not just errors; they are a routing decision. A retry policy can burn money quietly unless the loop records feedback, recovery attempts, and the point where escalation is cheaper than wandering.

Try this: For any coding-agent benchmark, log failed attempts separately: model used, feedback received, retry count, escalation point, and final cost per accepted fix.