AutoScientists-Codex

A Codex-native adaptation layer for mims-harvard/AutoScientists: install a pinned local runtime, launch or resume research runs, and execute the upstream multi-agent protocol with Codex-native subagents.

This is an independent compatibility project, not the Harvard authors' repository and not an official OpenAI project.

Why an adapter instead of a source fork?

The upstream public commit targeted by this release does not contain a repository-level open-source license. This repository therefore does not copy or relicense upstream code. Its installer fetches the pinned upstream commit directly from the authors' repository into a local, ignored runtime directory. Apache-2.0 covers only the independently written adapter files here.

See THIRD_PARTY_NOTICES.md for the exact boundary.

Codex adaptations

• A distributable Codex plugin and reusable $autoscientists-codex skill.
• Idempotent installation of a pinned upstream checkout, isolated Python dependencies, and pinned ClawInstitute runtime.
• Diagnostics for Git, Python, Node.js, Codex CLI, upstream revision, runtime dependencies, token state, and the coordination server.
• Codex-native mapping for upstream Agent(...) and Task(...) pseudocode: native spawn, wait, close, and fresh-session semantics.
• Generated CODEX_RUN.md and AGENTS.md overlays for every run without modifying the upstream template.
• Start, stop, task discovery, launch, resume preparation, prompt generation, and a codex exec fallback for surfaces without native subagents.
• No invented mapping from Claude model labels to Codex model IDs; Codex agents inherit the user's configured model unless explicitly overridden.

Install with a Codex prompt

Prerequisites: Codex CLI, Git, Python 3.9 or newer, Node.js 22 or newer, and npm.

Copy and paste this prompt into Codex:

Install AutoScientists-Codex from https://github.com/yptang98/AutoScientists-Codex.
Add yptang98/AutoScientists-Codex as a Codex plugin marketplace, install and
enable autoscientists-codex@autoscientists-codex, then use the installed
adapter to install its pinned local runtime, start ClawInstitute, and run strict
diagnostics. Verify that the plugin version is 0.1.2, the upstream checkout is
at the pinned commit, and runtime readiness has no errors or warnings. Do not
create a research run, dispatch agents, or start training. If Codex needs a new
task to load the installed skill, finish installation and tell me to open one.

After installation, open a new Codex task and paste:

Use $autoscientists-codex to verify runtime readiness and list the available
task families. Do not start a run yet.

Safety

AutoScientists can launch long-running experiments that consume substantial compute. Review the task's data license, resource policy, GPU mapping, protected paths, evaluation command, budget, and exit condition before starting a full run. The adapter does not upload private data.

Attribution

This adapter is based on the public interface and research design described by:

Shanghua Gao, Ada Fang, and Marinka Zitnik. “AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation.” arXiv:2605.28655 (2026). https://arxiv.org/abs/2605.28655

If you use the system in research, cite the original paper and identify this repository as the Codex compatibility layer. Machine-readable metadata is in CITATION.cff.

License

Apache License 2.0 for this adapter only. Upstream and third-party components retain their own terms; see NOTICE and THIRD_PARTY_NOTICES.md.