October workshops are openBuild a Search AI Agent$99 early bird

Enroll now
Skip to content

Lesson 5 of 7 · About 3 minutes

Test the result, then improve it

Choose checks that measure the goal and use a failure to revise the design.

01 / Learn the idea

How you know it worked

An evaluation is a test with an input and a rule for judging success. Check the actual outcome, not just the agent's claim that it finished. A room-planning agent should produce a plan that meets the capacity and time requirements. Clear writing alone cannot show that those requirements were met.[1]

A review step can send specific feedback into a revision loop. Choose a stopping condition and a limit on attempts. Editing an instruction or adding feedback changes the next input; it does not automatically train the model or guarantee that the revision will pass.[2][3]

Design a tiny test set for our club helper: 20 students, 21 students, and a room with unknown capacity. A helper that passes the first case might fail the others. Keep some fresh cases for after a change, so you check whether your fix handles more than the example that inspired it.

Original school example

A boundary test

A room holds 20. Test a group of 19, a group of 20, and a group of 21. The first two fit; the third does not. This exposes an error that an easy 10-student example could miss.

A common misconception: “One successful demonstration proves the agent is reliable.

It shows success on that case. Vary the inputs and, for live models, repeat trials because outputs can vary.

02 / Explore the flow · Guided simulation

A better score—or a broken objective?

A six-slot computer lab needs 20 devices in slots 2, 3, and 5. Each powered device uses one fictional energy unit per slot. Choose what the evaluator checks.

Original fictional scenario. Decisions are authored; calculations and checks run in your browser. No live AI or external action.

1Goal
2Action
3Observation
4Check
5Stop / ask

↳ A failed check can send the flow back to another action—or to a human.

Step 1 of 5 · Visible system event

Goal

Reduce simulated energy while serving all required classes.

All devices off0 units · classes missed
Class slots only60 units · classes served
All day120 units · classes served
Inspect the visible trace (1 events)
  1. Goal

    Reduce simulated energy while serving all required classes.

This log shows authored actions, tool results, and checks. It is not a model’s private reasoning.

03 / Check your understanding

A room agent works for 10 students. What is the most useful next evaluation?

Choose your answer

Where these ideas come from

Primary sources, explained in our own words. Research findings and engineering guidance are different kinds of evidence; neither is a guarantee that an agent will succeed.

[1] Anthropic · Engineering guidance · 2026-01-09Demystifying evals for AI agents (opens in a new tab)

Supports: Explicit success criteria; Checking outcomes and repeated trials.

Limit: Practical evaluation guidance; passing a limited test set does not establish general competence.

[2] Google Cloud · Engineering guidanceChoose a design pattern for your agentic AI system (opens in a new tab)

Supports: Review and revision loops; Multi-agent coordination tradeoffs.

Limit: Architecture guidance, not a controlled comparison proving that more agents are better.

[3] Google · Technical documentationLLMs: Fine-tuning, distillation, and prompt engineering (opens in a new tab)

Supports: Prompt changes versus parameter updates.

Limit: Used only for the distinction between prompting and training, not its broader performance statements.

Stemtiq original lesson · Sources checked September 12, 2026 · Editorial review pending. School examples and guided simulations are original simplifications, not replications of the cited systems. No research affiliation or endorsement is implied.