Course contents · How AI agents actually work
- 1What is an LLM?
- 2Prompts: brief it like a contractor
- 3Tools: how an AI does things
- 4Context: what's on the desk
- 5Memory: the filing cabinet
- 6What is an AI agent?
- 7Evals: how do you know it works?
- 8Who approves what an agent does?
- 9RAG: finding the right page
- 10Subagents: when an agent delegates
- 11MCP: one plug for every tool
- 12Not everything needs to be an agent
- 13The whole picture
Lesson 7 of 13
Evals: how do you know it works?
Why "I tried it a few times and it seemed fine" isn't enough, and the four layers of checks that tell you an agent actually works.
CTO and co-founder, LaunchLemonade8 minReading timeAsk most people how they know their AI agent works and the honest answer is: "I tried it a few times and it seemed fine." That's fine for a demo. It isn't fine for a job you hand over every day.
It passed once
An LLM is not deterministic. The same input can produce different outputs. So when your agent handles a test case well, it didn't pass. It passed once.
Run it fifty times a day for real customers and "seemed fine" starts to wobble. Evals are how you measure an AI system instead of vibe-checking it.
- Testing asks
- does the code do what I wrote?
- Evals ask
- does the system do what I meant?
The idea comes from normal software testing. But a classic test checks an exact answer: 2 + 2 must equal 4, every time. Agent outputs are fuzzy. An email can be worded a hundred ways and still be a good email. So evals come in layers.
The four layers
- 01Hard checks
- Did it call the right tool? Is the output valid? Did it stay under budget? Did it stop? Cheap and unambiguous, so use them everywhere.
- 02Golden sets
- A fixed list of real tasks with known-good answers. Run the whole set after every change. If the score drops, the change made things worse, however clever it felt.
- 03LLM as judge
- One model grades another's output against a written rubric. Imperfect, but it scales in a way people can't.
- 04Human review
- Still the ground truth. Just too expensive to be the only truth.
Each layer catches what the one before can't. Hard checks catch broken mechanics. Golden sets catch regressions. A judge model catches tone and quality at scale. People catch what all three missed, and decide what "good" means in the first place.
Every bug becomes an eval
The most useful habit is simple. When the agent gets something wrong for a real user, that case joins the test set. From then on, if that mistake comes back, the test set catches it before a client does.
- A user hits a failure
- Write it down as a test case with the right answer
- Add it to the golden set
- Every future change is checked against it
At LaunchLemonade we learned this the hard way. A model change would feel like an upgrade, then the golden set would come back several points worse. Without the numbers, we would have shipped it.
Your most valuable asset
Over time, an eval set becomes one of the most valuable things you own. It is the answer key for your product, built from your real work, and nobody else has it.
It also makes change safe. When a new model comes out, you don't have to guess whether to switch. You run the set and look at the score.
Running example
The running job
Chasing last month's unpaid invoices, one lesson at a time.
Before Priya trusts the invoice chaser with real clients, it gets a small golden set built from cases she knows matter:
- A client who paid yesterday must not be chased.
- A client disputing an invoice must not be chased.
- A client with two overdue invoices gets one email that mentions both.
- Every reminder quotes the right invoice number and amount.
The amounts and invoice numbers are hard checks. Tone (polite, firm, sounds like Priya) is graded by a judge model against a short rubric. And Priya still reads every draft before it goes, which makes her the last layer. When she catches something, it becomes a new case in the set.
Practice
Try it yourself
About five minutes, in any AI chat app
Build a three-case golden set
Pick a prompt you use often in any AI chat app, such as summarising meeting notes.
Write three test inputs with what a good answer must include or avoid. Make one of them awkward on purpose.
Run each input three times in fresh chats. Score each answer against your list.
What to look for
Answers that pass one run and fail the next. That variation is exactly why you test many times, and why you keep the cases for every future change.
Common mistakes
Testing once and calling it done.
Run each case several times. A single pass tells you very little about a system that varies.
Changing the model or prompt without re-running anything.
Keep a golden set and run it after every change, however small.
Fixing a bug and moving on.
Turn every real failure into a test case, so you find out if it comes back.
Quick check
3 questions to see what stuck. Nothing is scored or saved.
1. Why isn't one successful test enough for an agent?
2. What is a golden set?
3. A customer hits a new kind of failure. What's the habit?
Remember
Evals measure what your agent does instead of vibe-checking it. Every bug becomes a test case, and the set becomes your answer key.
Sources and further reading
- 1.The LinkedIn post this lesson grew out of: Evals
Finished this lesson?
0 of 13 done in this course. Progress stays in this browser, no account needed.