Stage your agent eval harness L1→L3 and score trajectories, not just final outputs
Agentic Development Handbook·medium signal
Mature agent eval harnesses progress through levels: L1 runs existing linters/tests against agent output, L2 auto-generates tests from acceptance criteria, and L3 adds LLM-as-a-Judge — combined into one quality gate before a human reviews anything. Critically, evaluate the trajectory, not just the answer: two agents can produce identical outputs while one read 3 files and the other read 30, with very different cost, latency, and failure modes. The teams shipping agents reliably in 2026 win on evaluation infrastructure, not on model choice.