# Optimize a recipe

Ask a coding agent to improve an existing recipe against its eval suite. The coding agent tests several changes and keeps the best result.

Recipe optimization can improve instructions, skills, tool rules, and model
settings.

You can also use optimization to make a cheaper or faster model perform more
like a stronger model. People sometimes call this **harness distillation**.
The model stays fixed while the recipe changes.

## Prerequisite

Install the [Introspection CLI](/sdk/cli#install) before you use this guide.

## Ask the coding agent to optimize the recipe

Start with a short request:

```text
Use Introspection to optimize this recipe.
```

Add an outcome when you have one. For example:

```text
Use Introspection to optimize this recipe to use GLM 5.3 and match the
performance of Opus 4.6 on our existing evals.
```

You do not need to choose the search settings or tell the coding agent which
files to change. The coding agent inspects the recipe, eval suite, saved
results, and repository commands before it starts.

If you know that a complete reference result already exists, you can mention
it to avoid another reference run:

```text
There is already a complete Opus result in the eval history. Reuse it if it is
still comparable, and do not run Opus again.
```

You can leave that detail out when you do not know. The coding agent decides
whether a saved result is suitable.

You can also set a budget when you want to limit or extend the search:

```text
Use up to two rounds and four candidates.
```

If you leave out the budget, the coding agent uses the workflow's default
limit.

## What happens next

The coding agent first reports what it found and how it will measure the
result. It asks a question only when it cannot infer an important part of the
goal. If the repository does not have a suitable eval suite, it explains what
is missing before it starts the search.

The coding agent then runs a limited search over several recipe changes. It
uses the existing eval suite to compare them and applies the best change that
improves the result. If no change improves the result, it leaves the recipe
unchanged.

Optimization changes local recipe files only. It does not deploy a runtime or
change live resources.

## Read the result

The final report tells you what the unchanged recipe scored, what the selected
recipe scored, and whether it met the target. It also names the selected Git
commit and the recipe changes that produced the result.

For example:

```text
Opus reference:      43/52
GLM baseline:        33/52
Selected GLM recipe: 43/52
Improvement:          10 more checks passed
Target:               met
```

If the search stops before it meets the target, you can keep the improvement or
ask the coding agent to try another limited search. You can also give the next
search a larger budget.

## Verify and ship the change

After the coding agent applies the selected change, try the agent locally:

```bash
introspection local
```

Review and commit the selected change. Then follow the
[Agent development lifecycle](/guides/development-lifecycle) to test the same
commit in development and staging before production.

## Next

- [Evaluate offline](/guides/evaluate-offline) explains how to create an eval
  suite when the repository does not have one.
- [Agent development lifecycle](/guides/development-lifecycle) explains how to
  test and ship the selected recipe.

Optimize a recipe

Ask a coding agent to improve an existing recipe against its eval suite. The coding agent tests several changes and keeps the best result.

Recipe optimization can improve instructions, skills, tool rules, and model settings.

You can also use optimization to make a cheaper or faster model perform more like a stronger model. People sometimes call this harness distillation. The model stays fixed while the recipe changes.

Prerequisite

Install the Introspection CLI before you use this guide.

Ask the coding agent to optimize the recipe

Start with a short request:


Use Introspection to optimize this recipe.

Add an outcome when you have one. For example:


Use Introspection to optimize this recipe to use GLM 5.3 and match the
performance of Opus 4.6 on our existing evals.

You do not need to choose the search settings or tell the coding agent which files to change. The coding agent inspects the recipe, eval suite, saved results, and repository commands before it starts.

If you know that a complete reference result already exists, you can mention it to avoid another reference run:


There is already a complete Opus result in the eval history. Reuse it if it is
still comparable, and do not run Opus again.

You can leave that detail out when you do not know. The coding agent decides whether a saved result is suitable.

You can also set a budget when you want to limit or extend the search:


Use up to two rounds and four candidates.

If you leave out the budget, the coding agent uses the workflow’s default limit.

What happens next

The coding agent first reports what it found and how it will measure the result. It asks a question only when it cannot infer an important part of the goal. If the repository does not have a suitable eval suite, it explains what is missing before it starts the search.

The coding agent then runs a limited search over several recipe changes. It uses the existing eval suite to compare them and applies the best change that improves the result. If no change improves the result, it leaves the recipe unchanged.

Optimization changes local recipe files only. It does not deploy a runtime or change live resources.

Read the result

The final report tells you what the unchanged recipe scored, what the selected recipe scored, and whether it met the target. It also names the selected Git commit and the recipe changes that produced the result.

For example:


Opus reference:      43/52
GLM baseline:        33/52
Selected GLM recipe: 43/52
Improvement:          10 more checks passed
Target:               met

If the search stops before it meets the target, you can keep the improvement or ask the coding agent to try another limited search. You can also give the next search a larger budget.

Verify and ship the change

After the coding agent applies the selected change, try the agent locally:


introspection local

Review and commit the selected change. Then follow the Agent development lifecycle to test the same commit in development and staging before production.

Evaluate offline explains how to create an eval suite when the repository does not have one.
Agent development lifecycle explains how to test and ship the selected recipe.