AgentLens Benchmark for Interactive Coding Agent Trajectories
July 6, 2026
AgentLens moves beyond binary pass/fail metrics to evaluate the full execution trajectories of interactive coding agents. It uses a combination of formal verification and LLM-written reviews to diagnose instruction following, tool usage, and error recovery behaviors.
HOW THIS AFFECTS YOU
●
builderYou can use these trajectory reviews to diagnose specific failure modes in your coding agents.
●
researcherThis provides a more granular evaluation framework for agentic reasoning and error recovery.