Skip to content
SciTechBiz Institute
I want to join
Blog · Forward Deployed Engineer

Evals for AI agents: measuring before you deploy

Evals for AI agents are tests on real cases that the agent passes before anyone uses it. The first suite is built from 20 to 50 tasks drawn from genuine failures. It separates what the agent cannot yet do from what it must never get wrong again, and that result decides whether it goes into production.

Updated: 5 November 2026 · Deibson Silva · Forward Deployed AI Engineer and Founder

What they are

What are evals for an AI agent?

An eval is a test with one specific case and a way of telling whether the agent got it right. A task is that case: the input the agent is given and what counts as a correct answer. A grader is whoever scores it. A suite is the set of tasks.

A demo shows that the agent can get it right once, with the example chosen by whoever prepared it. Evals say which cases it gets right, which it gets wrong and whether it fails always or only sometimes.

You need to measure more than once because an AI agent does not answer the same way on each attempt. The same task can go well in the morning and badly in the afternoon. So each task is run several times and the full transcript of each attempt is kept, with the steps the agent took.

In the method, the evals have an owner from the first day: a named person who maintains the suite and reads the results.

Comparison

How do capability evals differ from regression evals?

They are two separate suites and they are read differently.

Capability evalsRegression evals
What they measureWhat the agent cannot yet do wellWhat the agent already did well and must not get wrong again
Where the tasks come fromThe hard cases in the process and what you want it to reachReal failures already corrected and the cases the expert signed off
Result at the startLow: if everything passes, the suite is too easyHigh: almost everything has to pass
When they runWhile you build, to see whether a change helpsBefore any new version is released
If they failYou keep workingThe version does not go out
Step by step

How do you build a first suite of evals?

This is how it is done in the «Build with evals» stage of the Legathum AI-First Hybridisation Method v3:

  1. 1Name the owner of the evals. Without that person, the suite stops being maintained the moment the project ends.
  2. 2Collect real failures from today's process. In an accountancy practice that sorts its clients' documents: the crooked photo, the duplicated document, the missing VAT number.
  3. 3Write 20 to 50 tasks from those failures. Each one carries the input exactly as it arrives in real work and what counts as a pass. «Similar» cases, made up to fill the list, stay out.
  4. 4Add the cases the in-house expert solved when their criterion was captured. Part of them stays sealed, unseen by the agent and by whoever builds it, for the blind test at the end.
  5. 5Choose a grader for each task: code when the answer can be checked, a model with a rubric when it cannot, and a person who calibrates both.
  6. 6Run each task several times and read the transcripts. The number alone does not explain why it failed.
  7. 7Separate capability from regression. Once the agent is in use, every failure that is noted goes into the regression suite as a new task.

In the programme's labs the tasks use synthetic data only.

How to read them

What do pass@k and pass^k mean?

They are two ways of summing up several attempts at the same task. They say different things.

pass@k

The agent gets it right at least once in k attempts. Useful when a person will review the answer and can ask for another attempt.

pass^k

The agent gets it right on all k attempts. This is the measure that matters when it will do the task with nobody looking at each answer.

The gap between the two

An agent with a high pass@k and a low pass^k can do the task, but not every time. It is not yet fit to work alone.

The transcript

Two failures with the same score can have different causes: a badly written rule, a missing piece of data or a case out of scope. You only see that by reading.

Before release

Which tests does an agent pass before it goes into production?

Two, and the client signs both. The first is called Mirror: the in-house expert judges, blind, the agent's answers to the sealed cases. It passes with at least 90% agreement and no breach of a «never» in the Book of Criteria.

The second is the Security Test. The agent is attacked on purpose, with instructions hidden in messages and in documents. Each tool it uses is classified by what it can do, and a high-risk action needs a person's approval beforehand.

The agent's EU AI Act sheet is completed in that same test. Article 26 requires deployers to assign human oversight of a high-risk system to people with the necessary competence, training and authority. And Article 50 has applied since 2 August 2026: people talking to an AI system must be told it is an AI, unless that is obvious from the context.

No agent leaves shadow mode without passing the Security Test. After that it climbs the Autonomy Ladder one rung at a time: Shadow, Suggests, Approval, Sampling and Autonomous. It starts one rung below the target.

Afterwards

Who owns the evals when the project ends?

The client. At handover, someone on their team runs the regression suite from their own systems, reads the result and decides, without the person who built the agent.

If nobody in the company maintains the evals, the system stays as it was on the day it was delivered. The process changes, new cases arrive and nobody knows whether the agent is still getting things right.

That is why the suite travels with every version. The code, the instructions and the model version are released together, and the whole regression suite is run before release.

Good to know

Honest limits

  • A suite that passes only answers for the cases somebody thought to measure. The agent can still fail on others.
  • The thresholds in this article are those of the method taught in the certification. Each company sets its own with the owner of the process.
  • The programme's labs use synthetic data only. In the real project, the company's data never leaves its systems.
  • The evals lab and the security lab are the two you must pass to be certified.
  • Today the programme is taught in Spanish: 27 Spanish videos are published at the Academy, and the English videos arrive in November 2026.
  • The certification attests a competence, never a job or an income.
FAQ

What people ask most

01

What is an eval in artificial intelligence?

A test with one specific case and a way of telling whether the system got it right. For an agent it is run several times, because the agent does not answer the same way on each attempt, and the transcript is kept so each failure can be understood.

02

How many tasks do you need to start?

The method starts with 20 to 50 tasks drawn from real failures in the process, plus the cases the in-house expert solved when their criterion was captured. A few real tasks are worth more than many invented ones.

03

How is an eval different from a software test?

A software test always gives the same result for the same input. An AI agent can get it right on one attempt and wrong on the next. So evals repeat each task, use graders with a rubric when there is no single answer, and are read as percentages.

04

Who grades the agent's answers?

It depends on the task. Code when the answer can be checked, a model with a rubric when it cannot, and a person who calibrates both. In the final test the in-house expert judges blind.

05

When are the evals run again?

The whole regression suite is run before a new version is released, whether the change is to the code, the instructions or the model. And every failure that appears while the agent is in use becomes one more task.

Portrait of Deibson Silva
Deibson Silva · Forward Deployed AI Engineer and Founder

Founder of Legathum and iPersonality and Forward Deployed AI Engineer. He leads the SciTechBiz Community from Madrid and examines the FDE Certification defence.

Your licence

Start from your dashboard

On the licence page you request access; requesting access costs nothing. You then sign the contract and pay the annual licence from your dashboard.

iPersonality Global Institute, S.L. · NIF B27580042 · Madrid · Privacy policy

Prices in euros, exclusive of VAT. In Spain, Spanish VAT is added. In the rest of the EU, VAT is calculated at payment according to your country and VAT number. No Spanish VAT is charged to UK customers.