Building an Agent Eval Tool Reveals Why Scoring AI Agents Is Far Harder Than Models

A developer building AgentEval Forge, an open-source evaluation lab for AI agents, found that agent evaluation is fundamentally different from standard model evaluation. While model evaluation focuses on whether an output is correct, agent evaluation must assess the entire decision-making process, including tool selection, retries, cost, and recovery behavior. An agent can produce an acceptable final answer while taking inefficient, risky, or wasteful steps along the way, meaning final-output scoring alone can mask serious problems. The developer argues that the path an agent takes matters as much as — or more than — the endpoint, since cost, safety, and trust are determined during execution, not just at the result. This insight emerged not from research papers but from the practical challenges encountered while designing scenario packs, trajectory scoring, and regression tracking for real agent workflows.