RecoverRL

THE IDEA

A typical agent handles failure the same way every time — retry the API call, ask the LLM to reflect, or restart the whole task. RecoverRL learns to choose the right response from the failure itself, the execution history, and the current state of the environment.

THE SYSTEM

• —When a tool call fails, features are extracted from the error, the execution trace so far, and the environment state.
• —A learned recovery policy — supervised learning or a contextual bandit trained on execution traces — selects one action: retry, inspect state, refresh the schema, replan, or escalate.
• —Task completion is then verified, so a recovery only counts if the task actually finishes.

THE PROCESS

• —Built on the existing AgentGauntlet benchmark rather than a new failure-testing framework — the contribution is the learned policy and its results.
• —The policy is compared against fixed retries, rule-based recovery, and LLM self-reflection.
• —Evaluation tracks task recovery rate, cost per successful recovery, duplicate side effects, and performance on unfamiliar tools and task types.

WHERE IT STANDS

In development — the recovery policy and its comparison against baseline strategies are being built and evaluated now.