Full-Code · Orchestration Frameworksactive

CAMEL-AI

Type: full-code · Vendor: CAMEL-AI.org · Language: Python · License: Apache-2.0 · Status: active · Status in practice: mature

Links: homepage docs repo

Study agent scaling laws by providing a multi-agent framework whose core building blocks are ChatAgent (tool-calling LLM agent), RolePlaying (AI-assistant + AI-user dialectic), Workforce (managed multi-agent collaboration), MCP-backed toolkits, RAG pipelines, code interpreters, and large-scale simulated societies.

Description. CAMEL ('Communicative Agents for "Mind" Exploration of Large Language Model Society') is an Apache-2.0 Python multi-agent framework whose stated mission is 'finding the scaling laws of agents'. Building blocks: ChatAgent (atomic LLM-driven reasoning unit with tool use), RolePlaying (two-agent cooperative paradigm where an AI-assistant and an AI-user collaborate on a task), Workforce (manager-led multi-agent collaboration in camel/societies/workforce), 20+ toolkits including SearchToolkit, MCP support, RAG retrievers, code interpreters (Python/shell/browser), and OWL agent for reasoning. The framework is used at scale (up to ~1M agents in OASIS-style simulations).

Agent loop shape. ChatAgent inner loop is tool-calling: model emits messages and tool calls, tools (Python interpreter, shell, browser, MCP, search) execute, results return as messages. RolePlaying sits on top: a Society pairs an AI-assistant and an AI-user who exchange messages until the task converges or a step cap is hit. Workforce wraps this with a manager that decomposes tasks, assigns to workers, and synthesises results. Memory persists across turns; RAG pipelines combine chunking, retrieval, and generation; OASIS-style simulations scale this to large agent populations.

Primary use cases

two-agent role-playing task solving (CAMEL paradigm)
managed multi-agent collaboration via Workforce
ChatAgent-based tool-using assistants
MCP-backed agentic workflows over external infrastructure
large-scale social simulation (e.g. OASIS, ~1M agents)

flowchart TD USER[User task] --> RP[RolePlaying / Society] RP --> AIASSIST[AI-assistant<br/>ChatAgent] RP --> AIUSER[AI-user<br/>ChatAgent] AIASSIST <-->|messages| AIUSER AIASSIST -->|tool calls| TK[Toolkits] TK --> INTERP[Python / shell / browser interpreters] TK --> MCP[(MCP servers)] TK --> RAG[(RAG retrievers)] WF[Workforce<br/>manager + workers] --> AIASSIST WF --> WORKERS[Worker ChatAgents] AIASSIST -.memory.-> MEM[(Memory / storage)] WF -.OASIS.-> SIM[Large-scale simulation]

Key concepts

ChatAgent (docs) — Atomic reasoning unit driven by an LLM, capable of tool calls and decision-making.
RolePlaying / Society → camel-role-playing — Two-agent cooperative paradigm (AI-assistant + AI-user) for autonomous task solving.
Workforce → supervisor — Managed multi-agent collaboration in camel/societies/workforce (e.g. 'Hackathon Judge Committee with Workforce').
Toolkits + MCP → mcp — 20+ toolkits (search, web, code), with MCP support (ACI MCP, Cloudflare MCP CAMEL).
Interpreters → code-execution — Execution backends (Python, shell, browsers) for live code evaluation and automation.
RAG pipelines → agentic-rag — Combine chunking, retrieval, and generation for grounded responses.