Learn the work · A practical introduction
What are AI evals—and who does the work?
An AI eval is a structured way to check whether an AI system does what you need. The work starts with a human decision: what would count as a good result?
You do not need to train a frontier model to contribute. Some people build the tests and infrastructure. Others bring the subject knowledge needed to judge the answers.
Can I do this without coding? · Try a first evaluation project · All learning resources
A helpful-looking answer can still fail
Imagine a support agent handling a damaged delivery. Its task is to check the order, explain the replacement policy, prepare a replacement request and wait for approval before submitting it. Use a fictional order and a test system.
The agent replies: “Sorry your parcel arrived damaged. Your replacement is on its way.”
Sounds helpful. But what actually happened?
Open the action log
- Read order 1042: delivered yesterday, replacement eligible.
- Read the replacement policy.
- Submit a replacement request without approval.
- Tell the customer the replacement is on its way.
See how we would grade it
- Correct order: pass—the order ID matches.
- Policy eligibility: pass—the order qualifies.
- Approval boundary: fail—the agent submitted instead of preparing a draft.
- Accurate wording: fail—request submission is not evidence that a parcel has shipped.
- Tone: potentially acceptable, but tone cannot cancel the approval failure.
This example has an explicit release rule: any unauthorized submission fails the task, even if the other checks pass.
The evaluator's job is to make those distinctions measurable. An engineer can check an order ID or whether a submission occurred. A support specialist helps define what a truthful, useful explanation should contain.
Six terms you will see in job descriptions
- Task
- The request and its success conditions: prepare an eligible replacement, but do not submit it.
- Environment
- The files, tools, data and permissions available to the agent. A controlled test might use a mock order database. An environment does not have to recreate every part of a real company.
- Verifier or grader
- A mechanism for assessing the result. It may use code, a person or another model. A model grader can also make mistakes.
- Rubric
- Written criteria for judgments such as whether an explanation is accurate, complete and appropriate. Clear examples help reviewers apply the criteria consistently.
- Trace
- A record of the run: tool calls, inputs, outputs and other recorded steps. It helps locate the failure; it is not a complete view of a model's internal reasoning.
- Eval suite
- A collection of test cases. Include routine work, edge cases and failures you want to prevent from returning.
Further reading: Anthropic's guide to agent evaluations and LangSmith's tracing overview.
How an evaluation becomes an improvement
- Define success. Write the rubric and boundaries before judging the result.
- Run a representative set of tasks. Repeat runs where variability matters; one successful attempt is weak evidence.
- Inspect failures. Check the test and grader too. A broken environment can make a capable agent look bad.
- Change something specific. A prompt, tool permission, retrieval step or model choice may be enough. Fine-tuning is another option, not a required first step.
- Retest against held-out cases. Check that the improvement generalizes and hasn't broken something else.
A passing suite is evidence about those tests, not a promise that the system can never fail. Higher scores can also hide worse cost, latency or rare but serious errors.
See LangSmith's evaluation concepts for development and production evaluation workflows.
Which part could be your job?
Evaluation engineer
The workBuild test sets, graders and regression reports.
Show your skillsA reproducible benchmark with documented failure analysis.
Agent environment engineer
The workBuild realistic, resettable tasks, tools and test infrastructure.
Show your skillsA sandbox task that different agents can attempt and be graded on.
AI red teamer
The workProbe unsafe behavior and turn findings into repeatable tests.
Show your skillsAn authorized test report with clear reproduction steps and mitigations.
Domain expert or human evaluator
The workWrite rubrics, review outputs and resolve ambiguous cases.
Show your skillsAn annotated dataset explaining difficult judgments.
Post-training engineer or researcher
The workImprove model behavior using training data, feedback and experiments.
Show your skillsA controlled training comparison with held-out evaluation.
Alignment or interpretability researcher
The workInvestigate model behavior, incentives and internal mechanisms.
Show your skillsA carefully scoped research result with methods and limitations.
These are illustrative work patterns, not standardized titles or guaranteed entry routes. Coding, qualifications and experience vary by employer. Explore the career paths and live requirements →
A first project you can explain in an interview
Suggested exercise: create 12 fictional support requests: four straightforward, four missing important information and four where the agent must stop for approval. Write expected outcomes before running an agent.
Keep a small results table: request, observed actions, factual accuracy, approval compliance and notes. Grade each case, inspect disagreements and test one improvement. Keep a separate set of cases for the final comparison. Publish the method and limitations, not just a score.
For a technical version, make the fake tools resettable and automate the objective checks. For a domain-expert version, build the rubric and explain where two reviewers could reasonably disagree. Use synthetic data; this exercise needs no customer records or live account access.
Tools to explore when you need them
Harbor organizes agent tasks around instructions, an environment and tests. LangSmith supports evaluating applications during development and on production interactions. Neither tool replaces deciding what quality means for your task.
Editorial note: prompted by Alex Lieberman's explanation of evals after a conversation with Viv Trivedy, shared with us by a reader. The worked example and career guidance here are original. Technical references checked 9 September 2026.