# Autoresearch

C3 runs your research loop remotely. A research engine proposes changes to a program, evaluates them on C3 compute, and keeps the measured results. Submit from Python, close your notebook, and return to the same run when you want to inspect its progress.

**Start with [a short GPU experiment](https://docs.cthree.cloud/autoresearch/quickstart.md).** It includes a working candidate and evaluator, so you can see a complete run before adapting it to your own research.

## What a run contains[​](#what-a-run-contains "Direct link to What a run contains")

| Part            | What you choose or supply                                                              |
| --------------- | -------------------------------------------------------------------------------------- |
| Research engine | OpenEvolve is the first supported engine.                                              |
| Language models | A model or weighted model ensemble served through OpenRouter.                          |
| Compute         | A C3 hardware profile and, optionally, a particular compute provider.                  |
| Candidate       | The initial Python program that the engine will try to improve.                        |
| Evaluator       | A Python function that measures each candidate and returns a score.                    |
| Limits          | Campaign duration, evaluation duration, number of evaluations, and an optional budget. |

The engine coordinates the loop while C3 handles GPU provisioning, job execution, logs, artifacts, access control and credits. The dashboard records the selected engine, models and compute configuration alongside the run. OpenEvolve is an engine choice, not the name of the C3 product; additional engines are not yet implemented.

## Follow the experiment[​](#follow-the-experiment "Direct link to Follow the experiment")

Each submission returns a durable ID and a private `/squeue/ar_…` link. The **Autoresearch** section of the dashboard, directly above **Jobs**, lists your runs. A run's overview diagram shows its planned loop and current phase; the details show individual evaluations, scores, compute jobs and saved outputs.

* [Quickstart](https://docs.cthree.cloud/autoresearch/quickstart.md): run the supplied short GPU experiment.
* [Runs and results](https://docs.cthree.cloud/autoresearch/runs-and-results.md): reconnect, inspect progress, collect results and cancel work.
* [Configuration](https://docs.cthree.cloud/autoresearch/configuration.md): choose an engine, model, compute and limits.

## Current availability[​](#current-availability "Direct link to Current availability")

Autoresearch is a staging pilot. Use the supplied pilot source revision, a C3 staging API key and account credits. Its API is `https://test.api.cthree.cloud` and its dashboard is [the staging Autoresearch page](https://c3-research-stg.vercel.app/dashboard/autoresearch). The Python SDK is installed from that source revision; these instructions do not assume a public package release.

The pilot supports OpenEvolve 0.2.26, sequential evaluations and one active campaign across the deployment. A crashed controller does not automatically resume. Completed progress and artifacts remain available. Production CLI and MCP submission of research campaigns are not enabled.