Which LLM eval tools do practitioners use, and where do they fall short?

Hamel Husain names LangSmith, Arize and Braintrust as the vendors he meets most often. Langfuse and Promptfoo are two other common options. These tools run datasets, judges and experiments. Their docs and practitioner writing point to the same gap: no tool decides what “good” means for your product. You still need error analysis, an expert-written rubric and labeled examples.

Which LLM eval tools do practitioners use?

Hamel Husain, a practitioner who writes widely on evals, names the vendors he sees most often: LangSmith, Arize and Braintrust (Husain and Shankar, Evals FAQ, 2026-09-18). He says he has no favorite and that support quality and “vibes” matter more than features. Langfuse and Promptfoo are two other common options.

The table below sums up what each tool’s own documentation says it does. It is not a ranking.

Tool	What its docs say it does
LangSmith	Offline evals on curated datasets. Evaluators: human review, code rules, LLM-as-judge and pairwise comparison. Online evals on live traces.
Braintrust	Each eval has data, a task and evaluators. Evaluators can be autoevals, LLM-as-judge or custom code.
Langfuse	LLM-as-a-judge, human annotation queues, user feedback, custom scores and dataset experiments that can fail a CI build, and a guide to calibrate an LLM judge against human labels.
Arize Phoenix	Two evaluator types: LLM-based (judge model) and code (deterministic logic).
Promptfoo	Open-source CLI and library for evaluating and red-teaming LLM apps, including in CI/CD.

Sources for the table: LangSmith docs, Braintrust docs, Langfuse docs, Arize Phoenix docs, Promptfoo docs.

What do these tools do well?

They remove plumbing work. You get dataset storage, experiment runs, score tracking and a place to compare versions. Langfuse says its offline experiments can run in CI/CD and fail a build when a score breaks a threshold (Langfuse docs). LangSmith separates offline checks on curated data from online checks on production traces (LangSmith docs).

That plumbing matters. A regression gate in CI is hard to build and easy to neglect.

Where do LLM eval tools fall short?

The limits below come from tool docs and named practitioners. Each has its source.

1. They do not know what “good” means for you. Husain and Shankar warn that vendors promote off-the-shelf metrics. Their FAQ quotes Husain’s course: vendors’ off-the-shelf metrics “ensnare engineers into superfluous tasks” (Husain and Shankar, Evals FAQ, 2026-09-18). In his field guide, he describes teams that celebrated a 10% gain in a “helpfulness score” while users still struggled (Hamel Husain, Field Guide, 2025-03).
2. Online scoring has no ground truth, in any tool. Braintrust’s docs state this plainly: live requests have no expected answer, so online scoring relies on LLM-as-a-judge scorers (Braintrust docs). A judge you have not checked against expert labels is a guess.
3. Judge bias is only partly handled. LangSmith’s pairwise guide offers a randomize_order option. It also says position bias “should mainly be addressed via prompt engineering” (LangSmith docs).
4. Criteria are not fixed in advance. A study by Shreya Shankar and co-authors found that “users need criteria to grade outputs, but grading outputs helps users define criteria” (Shankar et al., arXiv, 2024-04). A tool can store a rubric. It cannot write one that fits your experts.
5. Generic annotation screens slow review. Husain and Shankar recommend a custom annotation tool and say teams with one “iterate ~10x faster”. They add that off-the-shelf tools can suit large annotation teams with strict access control. Even then, they say many teams find the overhead not worth it (Husain and Shankar, Evals FAQ, 2026-09-18). This is one practitioner’s view, not a measured result.
6. Dashboards can hide the real problem. Husain and Shankar write: “I have rarely found dashboards to be helpful.” They prefer short reports with custom charts (Husain and Shankar, Evals FAQ, 2026-09-18).

Langfuse documents a way to check a judge against human labels. The labels themselves still have to come from your experts.

How should you choose between tools?

Pick the tool last. Use this order.

1. Read a sample of real traces and write down the failure types. Husain calls error analysis “the single most valuable activity in AI development” (Hamel Husain, Field Guide, 2025-03).
2. Ask one domain expert to give a pass or fail verdict with a short written reason for each trace. See how to build an LLM judge.
3. Build a judge. Check it against the expert labels. Report true positive and true negative rates separately, because raw agreement can mislead when classes are imbalanced (Hamel Husain, LLM judge guide, 2024-10).
4. Add deterministic code checks wherever a rule can do the job. Husain and Shankar advise favoring “assertions or other deterministic checks” over judges in CI (Husain and Shankar, Evals FAQ, 2026-09-18).
5. Now pick the tool that fits your stack, your data rules and your team. Any of the five can run steps 3 and 4.

Should you buy a tool or hire a service?

A tool runs your evals. A service helps you decide what to evaluate and who labels it. Many teams need both. A tool is the right buy when your team already knows its failure types and has experts ready to label. A service is the right buy when it does not. Clarity’s page on the eval tool versus eval service lays out the difference. For the full method, see how to evaluate an AI agent and the agent evaluation framework guide.

FAQ

Is Braintrust better than LangSmith?

Their docs describe similar building blocks: datasets, tasks, evaluators and experiments. None of the sources cited here is a neutral head-to-head test, so this page does not rank them. Pick on stack fit, data rules and support quality.

Can Promptfoo replace a paid eval platform?

Its intro page describes an open-source CLI and library that runs in CI/CD and covers red teaming (Promptfoo docs). Whether that is enough depends on how many people need to review outputs and how much hosting you want to run yourself.

Do I need an LLM judge at all?

Not always. Code checks cover anything a rule can test. Use a judge for judgment calls, and check it against expert labels before you trust it.

What is the first thing to do before choosing an eval tool?

Read real traces and write down how the agent fails. Every source above, from Husain to Shankar, starts with the data, not the dashboard.

How Clarity can help

Clarity by Epistemic Me is an AI-native services firm and an official partner of Parlance Labs. Clarity offers evals consulting for AI agents, done-with-you or done-for-you: error analysis, expert rubrics, golden datasets, LLM-as-judge checked against expert labels and regression gates. Lynn Chyi, AI Experience Architect at Dayforce, said the work improved “our human expert alignment scores by nearly 60%” (Dayforce case study). Engagements start at $100K+, and scope is agreed in the proposal. See /evals/ and how to evaluate an AI agent.

Sources

• Hamel Husain and Shreya Shankar — Frequently Asked Questions about AI Evals (2026-09-18)
• Hamel Husain — A Field Guide to Rapidly Improving AI Products (2025-03-24)
• Hamel Husain — Creating a LLM-as-a-Judge That Drives Business Results (2024-10-29)
• Shankar et al. — Who Validates the Validators? (arXiv, 2024-04-18)
• LangChain — LangSmith evaluation docs (accessed 2026-10-30)
• LangChain — LangSmith pairwise evaluation docs (accessed 2026-10-30)
• Braintrust — Evals guide (accessed 2026-10-30)
• Langfuse — Evaluation overview (accessed 2026-10-30)
• Arize — Phoenix evaluators (accessed 2026-10-30)
• Promptfoo — Introduction (accessed 2026-10-30)