General Agent: A Self-Evolving, Synthetic Agent Environment
We are open-sourcing general-agent, a fully synthetic environment that evolves its own task corpus across 1,040 domains, 4,504 tasks, and more than 8,000 unique tools for training and evaluating capable agents.
General Agent: A Self-Evolving, Synthetic Agent Environment Training capable agents requires exposure to diverse tasks and tools throughout the whole post-training pipeline. Yet, agentic environments with exposure to 1000s of tools remain scarce in the open-source community. Today, we are open-sourcing a first version of the general-agent environment ( Environments Hub ) — a fully synthetic environment capable of growing its task corpus to be more diverse and challenging over time. It formulates synthetic task creation as a 2-player game between two agents: Synthesizer — An agent tasked to syn
saved by
related reading
- General Agent: A Self-Evolving, Synthetic Agent Environmentprimeintellect.ai
- Scaling Agentic RL: 365,000+ Environments for SWE, Terminal, and Searchprimeintellect.ai
- renderers: Token-Level Templating for Agentic RLprimeintellect.ai
- LLM Powered Autonomous Agents | Lil'Loglilianweng.github.io
- Senior SWE-Benchsenior-swe-bench.snorkel.ai
- Demystifying evals for AI agents \ Anthropicanthropic.com
- [2304.03442] Generative Agents: Interactive Simulacra of Human Behaviorarxiv.org
- A Taxonomy of RL Environments for LLM Agentsleehanchung.github.io
- Towards self-driving codebases · Cursorcursor.com
- FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale | alphaXivalphaxiv.org
- WebArena-Infinity: Generating Browser Environments with Verifiable Tasks at Scalewebarena.dev
- How DeepMind's Generally Capable Agents Were Trained — LessWronglesswrong.com