facebookresearch

facebookresearch / HyperAgents

Public

Self-referential self-improving agents that can optimize for any computable task

395
64
100% credibility
Found Mar 24, 2026 at 397 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

HyperAgents is a research framework for iteratively evolving AI agents that modify their own code to excel at diverse computational challenges like games, proofs, and reviews.

How It Works

1
📖 Discover HyperAgents

You find this exciting project from researchers at Meta that lets AI agents improve themselves on tough challenges like games and math proofs.

2
💻 Get everything ready

Download the files and prepare a safe space to play with evolving AI helpers.

3
🔑 Connect smart thinkers

Link popular AI services so your agents can brainstorm and create better ideas.

4
🐳 Launch the playground

Start a secure container where agents can safely experiment and change their own code.

5
🚀 Pick challenges and evolve

Choose tasks like robot walking or paper reviews, then watch agents automatically improve over generations.

6
📊 See the progress

Check charts and reports showing how much smarter your agents got on each challenge.

🎉 Supercharged agents ready

You now have powerful, self-improved AI helpers crushing tasks that were once impossible.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 397 to 395 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is HyperAgents?

HyperAgents is a Python framework for building self-referential, self-improving agents that iteratively optimize code for any computable task. You feed it domains like RL environments, code benchmarks, or math proofs, and it evolves agents via a meta-loop: evaluate performance, generate code edits, and repeat. Outputs include agent archives, progress plots, and ensemble scores—all sandboxed in Docker for safe execution of untrusted generated code.

Why is it gaining traction?

It stands out by tackling open-ended optimization across diverse tasks, beating baselines like DGM on multitask setups via archive-based parent selection and optional ensembling. Tied to hyperagents 2025 discussions and a hyperagents workshop, its arXiv paper draws eyes from agent researchers seeking beyond one-shot prompting. Developers notice reliable evals and transfer learning that compounds gains over generations.

Who should use this?

AI researchers testing agent evolution on custom computable tasks, RL engineers auto-tuning reward functions in envs like BabyAI or Crafter, or coders automating polyglot benchmarks and IMO-style proofs. Ideal for Python/Docker setups where you want self-improving agents without manual iteration.

Verdict

Worth forking on hyperagent GitHub if you're into self-improving agents—1.0% credibility score reflects 395 stars and research freshness, so docs are sparse and expect setup tweaks. Strong starter for 2025 agent experiments, but production users wait for more polish.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.