Agents18h ago
AgentLens Benchmarks Coding Agent Trajectories
AgentLens replaces binary pass-fail metrics with full trajectory reviews for interactive code agents. The benchmark pairs formal verification with LLM-written critiques to analyze how agents recover from errors and use tools. This approach allows developers to diagnose specific behavioral failures. It moves evaluation beyond simple rankings toward granular, production-ready debugging for coding agents.