Review of Mainstream Agent Evaluation Tools Including Promptfoo, DeepEval and LangSmith

On 2026-11-11 at 08:58, an article titled "Agent 评测工具链盘点：Promptfoo、DeepEval 与 LangSmith" appeared on Juejin, taking stock of three mainstream Agent evaluation tools: Promptfoo, DeepEval, and LangSmith. The piece continues a running series about building evaluations for RAG applications. Its framing is that the previous five articles had already handed developers a complete assessment methodology — what to measure (the test set), how to measure it (metrics), and who judges the results (the grader). At that point, though, the methodology still lived in documentation. The article positions itself as practical reading for developers who are assembling an evaluation stack for RAG and Agent applications.

One minute later, at 08:59 on the same day, a post with the same title went up on cnblogs, covering the same three tools. This version stays within the same series context but sharpens the problem: turning the accumulated method into an evaluation report a manager can read requires an execution pipeline, not just a notebook. The missing steps it names are concrete — how questions get dispatched in batch, how results get judged one at a time, and how scores get aggregated. It goes further than the first piece by stressing batch test runs, per-item judgment, and automated report generation.

Read together, the two posts frame Promptfoo, DeepEval, and LangSmith as the final piece of the series. The methodology from the earlier articles is treated as settled; what these articles add is the plumbing that moves it from documentation into a repeatable workflow of sending tests, judging outputs item by item, and producing reports automatically. No scores, benchmark results, or adoption figures for the three tools appear in either piece, so the current state of the discussion is limited to positioning them as the execution layer for evaluation — the point where test sets, metrics, and graders are actually put to work.