CAMEL
Python · Apache-2.0
CAMEL is an open-source framework for multi-agent collaboration, supporting role-play, task decomposition, and coordinated execution.
Overview
CAMEL (MBZUAI) leans into role-playing and emergent dialog. AutoGen (Microsoft) leans into programmable conversation flows and code execution. We compare role abstractions, dialog mechanics, observability, enterprise adoption, and the gotchas each one ships with.
Python · Apache-2.0
CAMEL is an open-source framework for multi-agent collaboration, supporting role-play, task decomposition, and coordinated execution.
Python · CC-BY-4.0
Microsoft AutoGen is a multi-agent conversation framework that lets you create multiple agents to collaborate through dialogue and solve complex tasks.
| Best for | CAMEL | AutoGen |
|---|---|---|
| Role abstraction | Roles (user / assistant / task specifier / critic) are first-class. Each Agent class pairs a system message with a role-specific response generator; the role-playing prompts read like they came out of the original paper. | Roles are just system messages with a name. No built-in taxonomy, which is more flexible but means you design the prompt templates yourself. |
| Conversation mechanics | Role-playing is the headline. CAMEL gives agents distinct role prompts (e.g. python programmer + stock trader) so the dialog reads like two specialists collaborating, not two generic assistants. The ChatAgent class wraps a single role with tool-use and step-by-step response generation. | Multiple conversation topologies ship out of the box — GroupChat, Swarm, Magentic-One. A code executor (Docker / Jupyter) is built in, so agents can run Python and rewrite themselves mid-conversation. |
| Observability | Dialog logs dump to stdout + JSON files by default. The ChatAgent step-by-step generator interface can plug into external trace platforms like Langfuse or Phoenix. | AutoGen Studio (Microsoft) gives you a visual debugger, and Azure AI Foundry integration is first-class. Local traces ship through an OpenTelemetry exporter to Langfuse or Phoenix. |
| Enterprise adoption | Citation count in academia is high, but production case studies are rare. MBZUAI used CAMEL inside SilkRoad to generate training data; direct commercial adoption is limited. | Microsoft backing + AutoGen Studio + Azure AI Foundry is the easy PoC path. Several large consultancies use AutoGen as the skeleton for their internal multi-agent tools. |
| Common gotchas | Long role-playing chains still burn tokens fast (each turn is a full LLM call). The role prompts lean toward English and need re-tuning for Chinese-first scenarios. | GroupChat with more than five agents gets flaky — the speaker selector starts picking odd agents and you end up writing a custom selector function. Code execution is unsafe by default; Docker isolation is mandatory. |
| Metric | CAMEL | AutoGen |
|---|---|---|
| Stars | 17.6k | 60.5k |
| Forks | 2.0k | 9.1k |
| Language | Python | Python |
| License | Apache-2.0 | CC-BY-4.0 |
| Last commit | August 14, 2026 | April 15, 2026 |
Choose based on your primary workflow, language ecosystem, and integration needs. Review each project's documentation and recent GitHub activity before adopting it in production.
Depends on the use case. For open-ended dialog, research assistants, or code generation, AutoGen GroupChat with the code executor is more battle-tested. For role-playing, simulation, training, or synthetic data generation, CAMEL role-playing abstractions map directly.
CAMEL has over a thousand citations and is maintained by the MBZUAI team — strong academic footprint. AutoGen bridges research and industry with Microsoft Research backing. Pick CAMEL for role emergence research; pick AutoGen for programmable conversation research.
Yes, with glue code. CAMEL ChatAgent exposes a step-by-step response generator; AutoGen GroupChat supports custom speaker selection. You can register a CAMEL agent as an AutoGen speaker, or vice versa. Community bridge projects exist; nothing official.
AutoGen. Microsoft ships stronger docs, tutorials, and AutoGen Studio, so a newcomer can run a multi-agent demo in 30 minutes. CAMEL role-playing prompts read like they came out of the original paper — you need to read the paper to tune them well.