Recurse docs / 17
Runnable examples
Use an example to study one complete Recurse agent before designing your own. Ask your coding agent to explain which parts match your task and which assumptions, checks, or tools must change before it copies or adapts anything.
An example demonstrates one method under stated conditions. Its measurements do not prove that the same method will work for another domain or dataset.
Choose an example
| Example | What it demonstrates | Important requirements | Main evidence |
|---|---|---|---|
| Tiny Tuner | A small deterministic optimization loop with independent validation. | Standard Recurse access and the example's locked dependencies. | Validation metrics and best-model.json. |
| Backpack 3D | Visual construction, critique, revision, and artifact preservation. | An Astra-enabled service with native image support and a photograph you can share. | Editable Python, GLB geometry, rendered comparisons, and attempt history. |
Choose the example whose loop and evidence are closest to the problem you want to solve, not merely the example with the most similar output format.
Review an example before running it
For each example, ask your coding agent to identify:
| Element | What to inspect |
|---|---|
| Problem | The outcome the Recurse agent is trying to produce. |
| Inputs | Caller values, required files, formats, units, and ranges. |
| Search choices | What the Recurse agent may change between candidates. |
| Action tools | Operations that construct or revise a candidate. |
| Verifiers | Checks, measurements, and diagnostics used to judge candidates. |
| Completion | Acceptance target, fallback behavior, and saved result. |
| Evidence | Structured output, artifacts, and run measurements. |
| Prerequisites | Dependencies, external services, credentials, and expected resources. |
| Limits | Claims the example's evaluation does not establish. |
Tiny Tuner
Tiny Tuner searches feature and training configurations for a small classifier. It demonstrates the smallest Recurse pattern: construct a candidate, measure it with an independent scorer, revise from the evidence, and preserve the best measured model.
Use it to understand the structure of a Recurse agent rather than as evidence for a production machine-learning method. Its score applies to its fixed example data and verifier.
Ask your coding agent to show you:
- How input values become the per-run task.
- Which tools construct candidates and which tool verifies them.
- How validation history reaches the next attempt.
- How the final result agrees with
best-model.json. - What happens when the target is not reached before the trial limit.
Source: Tiny Tuner on GitHub
Backpack 3D
Backpack 3D uses gpt-6-astra to write a Python modeling program from a backpack photograph. Recurse executes the program, exports GLB geometry, renders three views, and returns the comparison images to the Recurse agent for critique and revision.
Ask your coding agent to show you:
- How the reference photograph reaches the Recurse agent as native visual input.
- How candidate programs become geometry and rendered comparisons.
- How critiques, failed attempts, and incumbent replacements are recorded.
- How the selected Python program, GLB, previews, receipt, and history are preserved.
- Which checks establish executable geometry and which quality decisions remain subjective.
The example requires an Astra-enabled service with native image support and explicitly selects gpt-6-astra; it does not fall back to another model. Use only a photograph you are permitted to send to the service. Running its generated modeling programs locally executes them with your local account's permissions.
Before adapting it, distinguish independently checked requirements from subjective visual judgment. Replace example-specific choices and checks with requirements from your own domain instead of changing only the prompt.
Source: Backpack 3D on GitHub
Ask for an adaptation proposal
A useful instruction to your coding agent is:
Read the Recurse skill and this example. Explain its objective, search choices, action tools, verifiers, completion condition, allowance, artifacts, and known limits. Then propose the smallest changes needed for my task. Do not run or deploy anything until I approve the proposal.
After approval, keep the original example as a baseline and version your adapted prompt, tools, manifest, and lockfile together. Use Evaluate and improve to compare changes and Costs and limits to set the allowance.