← All Papers The Bottleneck Inception Prompting Live Dialogue Deep Dive Quiz
NeurIPS 2023 · Foundational Multi-Agent Paper

Communicative Agents for
"Mind" Exploration (CAMEL)

How role-playing and Inception Prompting unlock autonomous cooperation between LLM agents without human intervention or conversational drift.

Try Interactive Role-Play Explore Inception Prompting
Authors: Guohao Li, Hasan Abed Al Kader Hammoud, et al.
Institution: KAUST
Published: March 2023 · NeurIPS 2023
Core Contribution: Inception Prompting & Role-Playing Framework
The Multi-Agent Problem

Why Naive Autonomous Agents Fail in Conversation

Putting two conversational LLMs together in a free-form loop quickly results in degenerated interactions, endless pleasantries, or hallucinated task completion.

Naive Loop Pitfalls
Endless Politeness: Agents repeatedly thank each other ("You're welcome! Let me know if you need more.") without advancing the task.
Role Inversion & Drift: The assistant starts giving instructions while the user starts doing the work, muddying boundaries.
Premature Halting: One agent hallucinates that the software has already been compiled and deployed, stopping exploration.
The CAMEL Solution
Task Specifier: A preparatory LLM converts an abstract user idea (e.g., "build a game") into a concrete, actionable specification with strict constraints.
Inception Prompting: Asymmetric system messages with complementary rules (AI User gives instructions & evaluations; AI Assistant writes code & answers).
Strict Termination Protocols: Concrete token triggers ("") and step limits prevent infinite looping.
Framework Architecture

The Three Pillars of Inception Prompting

CAMEL designs static prompts applied at session initialization to steer communicative agents toward productive problem-solving.

Pillar 1

Task Specifier

Takes the user's vague prompt ("Develop a trading bot") and fleshes out specific indicators, APIs, stop-loss logic, and programming language before dialogue begins.

Pillar 2

AI User Agent

Plays the role of the client/planner. Bound by inception rules never to write solution code itself, but always to provide clear, single-step tasks and verify assistant deliverables.

Pillar 3

AI Assistant Agent

Plays the role of the technical domain expert. Bound never to ask generic open-ended questions, but to supply precise solutions followed by actionable next-step proposals.

Interactive Playground

Step Through an Autonomous Agent Society

Select a collaborative domain and click 'Next Turn' to observe how CAMEL agents coordinate step-by-step.

Deep Dive Lab

Why Role Assignment Scales Collaboration

Analysis of multi-agent dynamics: communication efficiency, hallucination suppression, and autonomous instruction generation.

Hallucination Mitigation in Societies

When a single LLM attempts end-to-end engineering, it frequently hallucinates intermediate progress because it holds both the goal and the implementation in one context.

The Critic Effect: By separating the User (evaluator) from the Assistant (coder), the User agent cross-examines outputs against specifications. If code lacks test cases, the User agent rejects it on the next turn.
Data Generation Scale

CAMEL was used to synthesize massive conversational instruction datasets:

  • AI Society Dataset: 25,000+ multi-turn dialogues across 50 roles and 1,000 tasks.
  • Code Dataset: 20,000+ problem-solution pairs with verifiable test scripts.
  • Zero Human Demonstrations: Generated entirely through autonomous role-play.
Test Yourself

CAMEL Mastery Quiz

Test your understanding of communicative multi-agent architectures and Inception Prompting.