Learn › Agent Harness, the big picture › Stop 7 › What a harness must do

Stop 7 of 9 · Pages

Prove and change: evals, tests and release

How you know the agent is good, and how you improve it without breaking it.

Requirement 25 of 33

Evals

Also called: Evaluations, golden sets, graders

Is it any good, and how would we know?

Real questionseach with itsright outcomeRun it and grade itby code, by a model,by a personA score"92 of 100 right"
An eval is a driving test for an agent.

Why

The same input can give a different output each time, so ordinary pass-or-fail tests are not enough. An eval runs many cases and reports how many came out right.

Without it

"It seemed fine when I tried it" is the only evidence.

Think about

  • A set of real cases with known good outcomes
  • How each is graded: a rule in code, a model as judge, or a person
  • When it runs: before a release, and on live traffic afterwards

In the big picture: Eval library, in the Change family

Requirement 26 of 33

Attack it before others do

Also called: Red teaming, adversarial testing

Does the harness hold when someone tries to break it?

Tricks andplanted ordersDoes it hold?run every attackyou knowA new testfor every holethat is found
The six stories in stop 5, run on purpose.

Why

Attackers will try hidden orders, false identities and persuasion. It is cheaper to try them first.

Without it

The first test of the guardrails is a real attack.

Think about

  • A library of known attacks, run before every release
  • People whose job is to find new ones
  • Every hole found becomes a permanent test

In the big picture: Eval library, and the Lifecycle rail

Requirement 27 of 33

Release safely

Also called: Versioning, staged rollout, rollback

How does a change reach users without hurting them?

A new versionprompt, modelor codeA small group firstwatch the numbersEveryoneor roll backin minutes
A changed prompt is a changed product.

Why

Changing one sentence in a prompt, or moving to a newer model, can change behaviour everywhere. Each change needs a version and a way back.

Without it

Every improvement is a gamble on live customers.

Think about

  • A version for prompts, models and tools, not only for code
  • The evals as a gate that every change must pass
  • A small group first, and a quick way back

In the big picture: Release and config, in the Change family

Requirement 28 of 33

Change settings without changing code

Also called: Configuration

Can we change a limit without a new release?

The ownerraises a limitSettings in one placeprompts, limits,the choice of modelThe agentpicks it up, nonew code
What changes often should not need a programmer.

Why

Limits, prompts and model choices change far more often than code. Keeping them as settings makes each change small, visible and reversible.

Without it

A one-line change to a limit waits two weeks for a release.

Think about

  • Which things are settings: limits, prompts, models, lists of allowed tools
  • Who may change each one
  • A record of every change, because settings are powerful

In the big picture: Release and config, in the Change family

Requirement 29 of 33

Learn from every miss

Also called: Feedback loop

How does a mistake in production make the agent better?

A missa complaint, a badanswer, an attackBecomes a casewith the rightoutcome written downIn the evalsrun before everyrelease, for ever
The same mistake should be impossible to make twice.

Why

Production finds the cases nobody imagined. Each one is worth keeping, because it is a real question with a known right outcome.

Without it

The same failure returns after the next change.

Think about

  • How misses are noticed: complaints, thumbs down, alerts, reviews
  • Who turns a miss into a case
  • How the set of cases is kept healthy as it grows

In the big picture: The Lifecycle rail, and the Eval library

← SeeProtect →