SKIP TO CONTENT
AI AGENTS6 minAgents in Production / Ep. 10 · Stage 10-1

AI Agent Tests: How to Set the Score Needed to Launch

Someone asks what score your AI agent needs before launch. If you answer with a copied number, your gate tests someone else’s risk.

XIN

IN 30 SECONDS

An eval threshold is the score an AI agent must reach on a fixed set of test cases before a new version ships. You earn your own: freeze the set, price each kind of mistake, write down why.

KEY TAKEAWAYS · 3

  • A threshold means something only next to its frozen eval set, its reason and the person who owns the risk.
  • Freeze the eval set before you measure: if you can edit the set, you can move the bar.
  • A copied threshold passes a gate that never tested your risk.
On this page

A copied threshold tests someone else’s risk

IN SHORT A threshold taken from a benchmark or another team gives a verdict on their deployment, not on yours.

WORDS TO KNOW

Threshold
The score a new version must reach on one measured dimension before it may ship. Also called the bar.
Eval set
The test cases you run the agent on, each labelled with the behavior you expect.
Frozen eval set
A released version of the eval set that never changes. Any edit makes a new version.
Baseline
The first score of the untouched system on the frozen set. Later runs are compared to it.

Episode 3 refused to publish an accuracy figure. Episode 6 built the release gate and left the bars to you. Episode 7 showed that, under its stated assumptions, escalated cases can dominate the bill.

I held back on purpose: a threshold from my deployment measures my risk, not yours. So here is the method instead, the part that transfers, limits of the evidence included.

WATCH OUT

The method comes from an 11-agent platform with one shared eval harness, the code that runs the tests. That count describes my documented work. Even the zero-failure example is a policy for known test cases, not a universal rule or a guarantee of safety.

COPIED VS EARNED

A copied threshold

A bar taken from a benchmark or another team. The gate still says pass or fail, but about somebody else’s deployment.

An earned threshold

A bar measured on your own frozen set, with its reason, owner and review date. It moves when your cost of being wrong moves.
Illustrative. No gate at all is an honest gap: everyone can see the evidence is missing.

A borrowed threshold looks like evidence, so it is worse than no gate at all.

Coming up next: So what does an earned threshold carry that a copied one doesn’t?

A threshold is a claim, not a number

IN SHORT A threshold is a record with seven fields, and the bar is the least interesting one.

IN PLAIN WORDS

An eval is an exam for your agent, run on an Eval set The test cases you run the agent on, each labelled with the behavior you expect. See the glossary. The Threshold The score a new version must reach on one measured dimension before it may ship. Also called the bar. See the glossary is the pass mark, except you write the exam, you set the mark, and you pay when the agent gets it wrong.

A number alone is only higher or lower than another. The six fields around it make it defensible.

THE THRESHOLD RECORD
record Threshold
dimensionstring

One threshold per scored dimension; one overall score is not a gate.

direction'at most' | 'at least'

Which way the bar points.

barnumber

The value: the only part anyone copies.

measured_onset_hash

The fingerprint of the frozen set it was earned on.

becausecost_of_being_wrong

Who pays when this dimension fails.

ownerperson

Who carries the risk, not who ran the eval.

review_bydate

When the bar must be re-earned, or it quietly becomes folklore.

A copied digit carries none of the other six fields.

The “because” field is the honest test: who pays when this dimension fails, and roughly how much? If you can’t say, you inherited the number.

ADAM says: Think of an exam: the frozen set is the paper, the threshold the pass mark. But two teams can rightly set different marks, and passing never proves no mistakes later.

Coming up next: So where does the bar’s value come from?

How a threshold gets earned, step by step

IN SHORT A threshold comes out of a loop you repeat on a frozen set until the expensive failures stop going down.

Freeze the test cases before the first run. A Frozen eval set A released version of the eval set that never changes. Any edit makes a new version. See the glossary never changes, because if you can edit the set, you can move the bar by editing the set. Adding or removing a case makes a new version, so only runs on the same version compare.

The first run of the untouched candidate (the version under test) is your Baseline The first score of the untouched system on the frozen set. Later runs are compared to it. See the glossary. Every later number is compared to it.

STEP BY STEP

Flow diagram: 5 stepsFlow diagram: 5 steps. Freeze the set, then Baseline run. Baseline run, then Sort failures by cost. Sort failures by cost, then Change one thing. Change one thing, then Re-run and compare. Re-run and compare loops back to Baseline run.1Freeze the set2Baseline run3Sort failures by cost4Change one thing5Re-run and compare

Step 1. Freeze. Note the set’s hash, a fingerprint of its content. Before that, no two runs compare.

Pick a step to see it move
  1. Freeze Note the set’s hash, a fingerprint of its content. Before that, no two runs compare.
  2. Baseline Run the candidate untouched: later runs compare to that score.
  3. Sort by cost Group failures by cost, not count. Ten cheap failures and one expensive one don’t weigh the same as eleven equal ones.
  4. Change one thing Change the prompt, retrieval, policy or model separately where practical. One better score doesn’t prove the cause: repeat runs to see how results vary.
  5. Re-run and compare Re-run on the same set version with the same labels, log every configuration change, compare, then back to step 2.
The dashed edge sends step 5 back to step 2: same set, one changed variable.

The loop shows where the system stands, not what is good enough. Pricing each way of being wrong decides that: a zero bar on known severe failures whose price nobody can name, a budget where it costs reviewer minutes.

ITEM GET

Earn the bar, don’t copy it

Your frozen set, your cost of being wrong, your risk owner: what a copied number lacks.

FAQ

Because it was earned on someone else’s data, failure modes and cost of being wrong. Copied into your gate, it judges a deployment you don’t have, and that verdict looks like evidence.

STAGE CLEAR

What you now know

  • What a threshold record holds besides the number: its set, its reason, its owner and a review date.
  • The loop: freeze, take a baseline, sort failures by cost, change one thing, re-run.

Check yourself

Tap or click a question to see the answer.

A vendor shows you the pass mark its agent cleared on a public benchmark. Can you make that mark your launch bar?

Not as it is: it came from their data and their cost of being wrong. Measure on your frozen set, with your own “because”.

Next stage:How do you turn the cost of each kind of mistake into the number itself?6 min
Adam Boudjemaa
Adam Boudjemaa

Web3 x AI engineer, former CTO of Integra. Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft.

Have a project to ship or a role to fill?

Discuss your project or role

My work at the web3 x AI intersectionOpen-source tools I build

LinkedInTwitterGitHub
Follow the engineering notes

Get new articles on production AI and smart contracts by email.