CAMEL

CAMEL is a Python framework for building multi-agent systems, oriented to research on the scaling laws of agents. It provides agents, tasks, prompts, model backends, tools, memory, and simulated environments, and it is used for three kinds of work: synthetic data generation, task automation through agent societies, and large-scale world simulation.

The core ChatAgent takes a model from ModelFactory and a list of tools, then answers step calls. Societies coordinate multiple agents: role-playing pairs and a Workforce that assigns tasks to a team. Data-generation modules implement chain-of-thought, Self-Instruct, Source2Synth, and self-improving pipelines, with cookbooks that fine-tune models on the output using Unsloth. Related research projects built on the framework include OWL, OASIS (social-media simulation), CRAB, and Loong, and the framework is designed for simulations of up to one million agents.

Agents can act as MCP clients, be exposed as MCP servers, and wrap toolkits as MCP servers. Optional logging writes model request and response JSON to a directory.

Features

• Agents: ChatAgent and other agent types with stateful memory for multi-step interaction
• Societies: role-playing agent pairs and Workforce teams for collaborative task completion
• Data generation: chain-of-thought, Self-Instruct, Source2Synth, and self-improving synthetic data pipelines
• Models: ModelFactory over multiple platforms, including OpenAI and local backends
• Toolkits: search, browser, terminal, and other tool integrations for agents
• Memory and storage: memory retrieval and persistent storage for agent state
• RAG components: loaders, embeddings, retrievers, and graph-RAG cookbooks
• Interpreters and runtimes: code execution and process management for agents
• Benchmarks: evaluation frameworks for agent performance
• MCP support: agents as MCP clients, agents and toolkits as MCP servers
• Human-in-the-loop: tool approval and human intervention components