---
name: recurse
description: Use when a task has a measurable or pass/fail target and would benefit from a dynamic sub-agent that can test and revise multiple attempts.
---

# Recurse

Recurse is a platform that lets an agent design dynamic sub-agents for complex workflows with
checkable outcomes. A dynamic sub-agent combines a goal, tools, and a code-based judge in one
deployed tool.

## Use a dynamic sub-agent when

- Success can be measured with a metric, test suite, scorer, or simulator.
- Each attempt should change in response to the previous measured result.
- The work needs repeated trials, a real environment, or more time than an inline task.

Handle one-shot checks, independent batches, and work without a checkable definition of done
directly.

## Working rule

Define the target and judge before starting. Pass relevant failed attempts and their scores as
seeds. Read the durable result and return it only when the stored outcome passed the judge. If
`returned_original` is true, return the original unchanged and say that no candidate passed.

## Connect

Install this file in the agent's skills directory and register the hosted MCP endpoint:

```sh
claude mcp add recurse https://mcp.recurse.run
```

Hosted MCP connections are unavailable during the private beta. See
[the documentation](https://recurse.run/docs) for SDK authoring and deployment guidance.