Agents1d ago
AgentLens Benchmarks Coding Agent Trajectories
The AgentLens benchmark moves beyond binary pass/fail metrics to evaluate the full execution path of coding agents. It combines formal verification with LLM-generated reviews to diagnose how agents use tools and recover from errors. This approach allows developers to pinpoint specific behavioral failures. Practitioners can now optimize agent trajectories rather than chasing single-score improvements.