Evidence-aware AI agent execution

I implemented an opt-in evidence-aware controller in the AssetOpsBench plan-execute agent. The controller combines argument validation, failure classification, replay restrictions, attempt limits, dependency checks, and outcome-dependent finalization.

Before a tool call, it checks advertised arguments, required fields, and basic types. After invocation, it recognizes exceptions, MCP error flags, empty responses, and application-level error payloads.

The extension preserves the default execution path. Recovery behavior is explicitly enabled rather than silently changing existing benchmark runs.

• Bounded attempts: at most one extra attempt per eligible step, within a default run-wide allowance of two.
• Replay restrictions: retries require explicit read-only metadata and no destructive flag. Mutating or unknown-safety calls are not replayed.
• Prerequisite enforcement: failed or missing dependencies block dependent steps before invocation; independent steps may continue.
• Explicit outcomes: exhausted recovery and failed dependencies route to deterministic failure reports.

I evaluated the production planner/parser, executor, and finalizer using seven fixture cases, two summary scripts, two conditions, and five repetitions: 140 deterministic offline executions.

In the permanent-prerequisite failure case, recovery exhausted one retry and blocked four dependent steps. The task still failed, but later actions did not proceed on missing evidence.

This work is documented in Execution Integrity through Bounded Recovery for Plan-Execute Agents, coauthored with Dhaval Patel.

The contribution is the integration of recovery boundaries into the executor, with inspectable traces and measured conditional behavior. Correct metadata and declared dependencies remain important assumptions.