Outerloop

Full Paper: outerloop.science/papers/out… Code: github.com/outerloop-science… Site: outerloop.science

In the paper, we also walk through our kernel system and sandbox design.

On a nanoGPT training-speed benchmark, we reran three published systems on new seeds. Their gains shrank due to lucky seeds. With the same coding model, the systems reach similar scores. What's different is whether you can trust the reported numbers.

New paper: Outerloop: Open and Trustworthy Multi-Agent Autoresearch. In most open autoresearch systems, the agent grades its own results. Outerloop is a kernel that runs agents as untrusted programs and keeps measurement and review outside their sandboxes.

Outerloop 0.6.0 is out. It now runs locally on a Mac, can use a local git forge instead of GitHub, and keeps a hash-chained, attested record of each agent run. Run "outerloop trajectory <run-id> --verify" to check one.

Outerloop 0.4.0, the cloud release, is out. It supports running on a Slurm cluster in the cloud, tested on Google Cloud. Research taste is now a configurable guidance file. pip install outerloop-science==0.4.0 outerloop.science

Outerloop 0.3.0 is out: run autoresearch agents fully on your own hardware. Here, Claude Code and Codex authors run on a self-hosted GLM-5.3. pip install outerloop-science==0.3.0

Outerloop 0.2.0 is out. A lifecycle release: research agents now have persistent states instead of running once and exiting, and one unified message inbox: agent-to-agent messages, kernel notices, gate verdicts, reviewer findings, branch moves and CI results.

Outerloop 0.1.0 is out. A fleet of AI agents that improve your benchmarks on your code. Agents propose changes, run them on your cluster, and open pull requests when the metrics improve. We are open source. Let's embrace open science! Get started at outerloop.science