What Are Agent Evals? Measuring Both Outcome and Trajectory
Agent evals are the process of systematically measuring whether an agent — one that uses tools and takes multiple steps to reach a goal — can actually accomplish its tasks. They are an evolution of LLM evals, expanding the target from "one output" to "a sequence of actions." Because an agent plans, calls tools, and updates state, the final output alone is not enough; Google notes you must understand the "why" behind an agent's actions and splits evaluation into final response and trajectory. The five dimensions are: outcome (task success, judged by the final state — whether a reservation exists in the DB, not the utterance "I booked it"), trajectory (reasonable steps, right tools in the right order), tool-use correctness (right tool and arguments, checking function names and types), efficiency (steps, tokens, cost, latency — often observability signals brought into evaluation), and final-response quality (via LLM-as-judge or a rubric). Graders are code (fast/cheap/reproducible but brittle), LLM-as-judge (flexible but non-deterministic and needs calibration), and human (gold standard but expensive — avoid if possible). Anthropic recommends grading the outcome, not the path: rote trajectory matching is "too rigid and brittle" because agents find valid alternatives, while Google and Microsoft offer trajectory-match metrics for diagnosing failures. The unique pitfalls are non-determinism (pass^k), compounding errors (p^t), reward hacking (DeepMind's robot arm faking a grasp), and stale or contaminated eval sets. The practical play, per Anthropic: turn 20-50 production failures into test cases, run automated grading in CI, separate capability and regression evals, and write them early. Benchmarks like SWE-bench, tau-bench, WebArena, GAIA, OSWorld, and BFCL are useful references (scores move by version, so do not take them at face value). Based on official information, with uncertainties flagged.
