camel

CAMEL is an open-source Python framework for building systems where many AI agents (programs powered by large language models) talk to each other to get things done, rather than relying on a single chatbot. It comes from a community research effort studying how agent societies behave as you scale them up — what the project calls finding the scaling laws of agents.

The framework is built around four design principles. Agents are evolvable, learning by interacting with environments and from data via reinforcement learning with verifiable rewards or supervised learning. They are scalable: CAMEL claims it can simulate up to one million agents at once. They are stateful, with each agent keeping a memory so it can carry out multi-step tasks. It also follows a code-as-prompt idea: code and comments themselves act as instructions agents read.

What you build with it falls into three buckets. The first is data generation: producing synthetic datasets with techniques the README names, such as Chain-of-Thought, Self-Instruct, Source2Synth, and Self-Improving CoT. The second is task automation, where agents take roles and cooperate on a job — role-playing societies, a Workforce module, and RAG pipelines that let agents look things up in your data are included. The third is world simulation, modelling environments populated by agents.

Reach for CAMEL if you are researching multi-agent systems, want to generate synthetic data to train or evaluate models, or want to build automation that needs several agents working together. It ships as a Python package with cookbooks, benchmarks, and an active community. The full README is longer than what was provided.