# Which agent benchmark should I use to test a coding agent?

For a bounded bug fix on a codebase the model hasn't seen, use SWE-bench Pro; for longer, multi-step work in a terminal, use Terminal-Bench; treat CursorBench as one extra signal rather than a primary score, since its tasks and grading method are both private.

The three aren't measuring the same thing. [SWE-bench Pro](/answers/swe-bench/what-is-swe-bench-pro) checks whether a model can land a working fix against a project's own tests, a single well-defined ticket. Terminal-Bench hands the agent a whole environment and a job that can take up to eight hours, closer to a day of real work than one patch. CursorBench draws its tasks from real requests Cursor users typed in its editor, which makes it the most realistic of the three and the least checkable, since nobody outside Cursor can verify a result and Cursor also sells a model scored on it.

None of the three predicts how your agent handles your own repository. [A benchmark score doesn't transfer to your setup](/answers/agent-reliability/how-do-you-measure-agent-reliability-in-production), because the prompts, retries, and tool wrappers around the model are the benchmark's, not yours.

---

Source: https://tessary.ai/answers/agent-benchmarks/which-agent-benchmark-should-i-use-for-a-coding-agent
More on Agent benchmarks: https://tessary.ai/answers/agent-benchmarks
From Tessary, agent reliability for AI agents in production: https://tessary.ai