CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents

What Happened

CodeRescue introduces a routing mechanism for coding agents that recovers from task failures within predefined compute budgets. Rather than restarting a failed run or abandoning the task, the system reroutes execution paths while tracking accumulated spend against a hard ceiling. The design treats recovery as a routing decision parameterized by remaining budget, not a static retry policy.

Why It Matters

Failed agent runs currently consume compute with no guaranteed output, and the standard fallback—full restart—compounds that waste multiplicatively across retries. For operators running autonomous coding pipelines, this converts an unpredictable cost tail into a bounded one. The benefit accrues most directly to teams running high-volume, long-horizon coding tasks where a single failure can invalidate expensive prior context. Budget-aware recovery also changes the economics of parallel agent deployment: if per-task spend is capped by construction, capacity planning becomes tractable rather than probabilistic.

Technical Details

The mechanism operates at the orchestration layer, intercepting failure signals and selecting an alternative execution path based on remaining budget, failure class, and prior trajectory. Recovery is budget-calibrated, meaning the router allocates a recovery allowance proportional to sunk cost and task state, rather than a fixed retry count. Integration requires the orchestration layer to expose per-task compute accounting and to surface structured failure metadata—raw exceptions are insufficient for routing decisions. The approach assumes failures are diagnosable enough to map to alternate paths; ambiguous or under-specified errors likely degrade to conventional retry behavior. No published benchmark numbers are available in the current material, so throughput and success-rate deltas remain unverified.

Operational Impact

Builders should wire recovery routing into the same layer that handles retries, timeouts, and circuit breaking, rather than as a separate fallback path. Day-to-day, this means per-task budget tracking becomes a first-class requirement—agents that cannot report spend cannot participate in routing. Manual intervention on common failure classes (dependency resolution, test failures, malformed diffs) should decline, shifting operator time toward failure-mode taxonomy and path design. Cost forecasting improves because worst-case retry overhead is replaced by a declared ceiling. The main workflow change: failure handling moves from exception management to policy configuration.

What To Watch

Expect budget governors to become standard alongside routing policies, mirroring timeout-aware patterns in distributed systems; the two are only useful together. The open question is whether recovery routing generalizes beyond coding to other agentic domains with cheaper failure signals. Second-order effect: if recovery becomes cheap and bounded, teams may raise task ambition per run, partially offsetting the projected savings.