ToolSuture: Recovering Agent Capability When APIs Change

Описание и применение

ToolSuture восстанавливает функциональность агента после изменения API, не меняя сам агент, используя Gemini для семантического анализа и проверку результата.

Задача

Agents break when underlying API contracts change, even if the agent logic remains correct.

Результат

ToolSuture restored the original shipment agent’s capability after a provider contract upgrade without modifying the agent code, achieving 0 bytes changed and 0 execution attempts in dangerous scenarios.

Рабочий процесс

• Freeze deployed agent and capture original mission
• Detect API change and obtain new tool contract
• Observe mission, old and new contracts, and runtime evidence
• Diagnose semantic equivalence using Gemini 3.6 Flash
• Apply policy to decide repair or refusal
• Plan bounded compatibility repair
• Validate repair against new contract
• Replay original mission with repaired environment
• Verify success through independent evidence checks

Доказательства

MISSION COMPLETED AND VERIFIED CAPABILITY_LOST → CAPABILITY_RESTORED 0 BYTES CHANGED

Источник

REFUSE CRITICAL BLOCKED SAFE_HOLD 0 EXECUTION ATTEMPTS

Источник