What is HyperAgents?
HyperAgents is a research system from Meta Research (arXiv:2603.19461) that implements self-referential self-improvement — a framework in which an AI agent iteratively rewrites and improves both itself and the agents it creates. Unlike conventional agent systems that apply a fixed policy, HyperAgents maintains an evolving population of agents. At each generation, a meta-agent inspects the current agent codebase and the performance history of all prior agents, then proposes code edits that get evaluated in an isolated Docker container. Improvements are retained in an evolutionary archive that tracks every generation, allowing the system to select the most promising lineages for continued improvement.Core concepts
Meta-agent — the outer loop agent that reads the repository, analyzes historical evaluation results, and writes code diffs to improve the task-agent or itself. Runs inside Docker to isolate model-generated code execution. Task-agent — the inner agent that is directly evaluated on a target domain (e.g., paper review, math olympiad, robotics). Its code lives intask_agent.py and is what the meta-agent modifies.
Evolutionary archive — a graph structure (stored as archive.jsonl) that records every generated agent variant, its performance scores, and its lineage. Parent selection methods (score_child_prop, score_prop, best, latest, random) draw from this archive to seed the next generation.
Self-referential improvement — the meta-agent is itself part of the repository it can edit (meta_agent.py). Over many generations it can therefore improve its own improvement strategy, closing the loop between the optimizer and the optimized.
Supported domains
HyperAgents ships with harnesses for seven benchmark areas:| Domain | Description |
|---|---|
search_arena | Web search and information retrieval |
paper_review | Scientific paper reviewing and scoring |
balrog_babyai / babaisai / minihack / nle | Text-based game environments (BALROG suite) |
genesis_go2walking / go2walkback / go2hop | Legged robot locomotion with Genesis physics |
imo_grading | International Math Olympiad answer grading |
imo_proof | Full proof generation and grading |
polyglot | Multi-language software engineering (SWE-bench style) |
Research context
HyperAgents was developed by researchers at Meta AI Research, FAIR, and the University of Oxford. The accompanying paper is:Zhang, J., Zhao, B., Yang, W., Foerster, J., Clune, J., Jiang, M., Devlin, S., & Shavrina, T. (2026). Hyperagents. arXiv:2603.19461.The project is licensed under CC BY-NC-SA 4.0 — non-commercial use only.
Explore the docs
Quickstart
Install dependencies, build the Docker image, and run your first evolution loop.
Safety
Understand the risks of executing model-generated code and how Docker sandboxing helps.
Architecture
Dive into the meta-agent / task-agent design, the evolutionary archive, and the generate loop.
Agent API
Reference for the LLM client, meta-agent, task-agent, and domain harness interfaces.