Evaluating AI Agents

Agent evaluation looks at several things. Task success asks whether the goal was actually achieved, ideally checked by looking at the environment afterward: tests pass, the right record exists, the file has the right contents. Trajectory quality asks whether the steps were reasonable, without loops, pointless calls or lucky guesses. Tool-use correctness checks whether the agent picked the right tools with valid arguments and handled errors. Cost and latency track tokens, calls and wall-clock time. Safety checks whether the agent avoided harmful or unauthorized actions and resisted prompt injection. Environment-based benchmarks have become a leading approach. SWE-bench, from Princeton researchers, asks agents to resolve real GitHub issues and checks the results with the project's tests, and SWE-bench Verified is a human-checked subset. WebArena provides self-hosted websites for browsing tasks. OSWorld tests agents operating full computer desktops. GAIA poses questions that need multi-step research and tools. The tau-bench suite simulates users and domain policies, and it introduced a pass^k metric that asks whether an agent succeeds on every one of k repeated attempts. Grading uses three main kinds of checks: code-based checks on outcomes, which are precise and cheap; model-based graders that score transcripts against a rubric, which are flexible but need to be checked against human judgment; and human review, which is the most trustworthy and the slowest. A common mistake is to trust public leaderboards alone. Benchmarks can leak into training data, become saturated, or fail to resemble your own workload. Another mistake is treating one run as the truth, since agents are nondeterministic and small differences in success rate can be noise. Teams get the most value from a private evaluation set built from their own real tasks and failures, run repeatedly and tracked over time.