HyperAgents: Meta Releases Self-Improving AI Agents That Rewrite Their Own Code
Meta's FAIR lab and Superintelligence Labs have open-sourced HyperAgents, a framework for building AI agents that can modify not just their task-solving code, but the code that governs how they improve — enabling metacognitive self-modification.
Unlike traditional agent frameworks where improvement loops are fixed by human engineers, HyperAgents collapses both the task agent and meta agent into a single editable Python program. The agent reads its own source code plus performance data, proposes targeted improvements via an LLM (Claude Sonnet), tests the new version against benchmarks, and keeps successful variants in an archive.
The research spans Meta FAIR, Meta Superintelligence Labs, University of British Columbia, and NYU. Results show cross-domain performance gains beyond coding tasks — the self-referential architecture enables agents to discover novel improvement strategies that fixed meta-agents cannot.
HyperAgents attracted over 1,000 GitHub stars in its first week and is currently trending on Hacker News with 147 points. The code is available at https://github.com/facebookresearch/HyperAgents under a Creative Commons non-commercial license.
This represents a significant step toward agents that don't just execute tasks but actively evolve their own reasoning and planning capabilities — a key building block for the self-improving agent paradigm.
← Back to all articles
Unlike traditional agent frameworks where improvement loops are fixed by human engineers, HyperAgents collapses both the task agent and meta agent into a single editable Python program. The agent reads its own source code plus performance data, proposes targeted improvements via an LLM (Claude Sonnet), tests the new version against benchmarks, and keeps successful variants in an archive.
The research spans Meta FAIR, Meta Superintelligence Labs, University of British Columbia, and NYU. Results show cross-domain performance gains beyond coding tasks — the self-referential architecture enables agents to discover novel improvement strategies that fixed meta-agents cannot.
HyperAgents attracted over 1,000 GitHub stars in its first week and is currently trending on Hacker News with 147 points. The code is available at https://github.com/facebookresearch/HyperAgents under a Creative Commons non-commercial license.
This represents a significant step toward agents that don't just execute tasks but actively evolve their own reasoning and planning capabilities — a key building block for the self-improving agent paradigm.
Comments