foundationagents/metagpt

🌟 The Multi-Agent Framework: First AI Software Company, Towards Natural Language Programming

Structural Verdict

How Data Flows Through the System

User requirements flow through specialized AI agents that collaborate to produce software artifacts, with each agent contributing domain expertise and passing results to the next stage.

1. Requirements Input — User provides natural language requirements or task descriptions
2. Product Planning — ProductManager analyzes requirements and creates product requirements document (config: llm.model, llm.api_type)
3. Architecture Design — Architect designs system structure, APIs, and data models (config: llm.model, llm.api_type)
4. Project Management — ProjectManager breaks down work into tasks and manages execution (config: llm.model)
5. Code Implementation — Engineers write code based on architecture and requirements (config: llm.model, llm.api_type)
6. Quality Assurance — QaEngineer tests and validates the implemented solution (config: llm.model)
7. Output Generation — Complete software project with documentation generated in workspace

System Behavior

How the system actually operates at runtime — where data accumulates, what loops, what waits, and what controls what.

Data Pools

Feedback Loops

• Code Review Loop (retry, balancing) — Trigger: Code quality issues detected. Action: Engineer revises code based on QA feedback. Exit: Code passes quality checks.
• Multi-Round Agent Collaboration (convergence, reinforcing) — Trigger: Team.run() called with n_round parameter. Action: Agents exchange messages and refine outputs. Exit: Max rounds reached or task complete.
• Workflow Optimization (training-loop, reinforcing) — Trigger: Optimizer starts evolutionary process. Action: Generate new workflow variants and test performance. Exit: Convergence criteria met.
• Android Learning Loop (training-loop, reinforcing) — Trigger: Learning stage activated. Action: Observe UI, take actions, refine operation docs. Exit: Task successfully learned or max attempts.

Delays & Async Processing

• LLM API Calls (async-processing, ~variable) — Agents wait for LLM responses before proceeding
• Minecraft Bot Ticks (rate-limit, ~50ms per tick) — Bot actions throttled to game tick rate
• Android Action Delays (async-processing, ~configurable) — Wait for UI changes after each action
• File I/O Operations (async-processing, ~variable) — Code generation and document creation delays

Control Points

• LLM Model Selection (env-var) — Controls: Which AI model powers agent reasoning. Default: gpt-4-turbo
• Investment Budget (runtime-toggle) — Controls: Maximum cost for LLM API usage. Default: 3.0
• Max Rounds (runtime-toggle) — Controls: Number of agent collaboration iterations. Default: 5
• Code Review Enabled (feature-flag) — Controls: Whether engineer uses code review process. Default: true
• Minecraft Wait Ticks (threshold) — Controls: Bot action timing and responsiveness

Technology Stack

Key Components

• Team (class) — Orchestrates multiple AI agents working together on projects metagpt/team.py
• Role (class) — Base class for all AI agent roles with action execution capabilities metagpt/roles/role.py
• Action (class) — Base class for all executable actions that agents can perform metagpt/actions/action.py
• Environment (class) — Base environment for agent interactions and state management metagpt/environment/base_env.py
• ProductManager (class) — AI agent that handles requirements analysis and product planning metagpt/roles/product_manager.py
• Architect (class) — AI agent that designs system architecture and APIs metagpt/roles/architect.py
• Engineer (class) — AI agent that writes and reviews code implementations metagpt/roles/engineer.py
• AndroidAssistant (class) — Specialized agent for learning and automating Android app interactions metagpt/ext/android_assistant/roles/android_assistant.py
• Optimizer (class) — Optimizes AI agent workflows using evolutionary algorithms metagpt/ext/aflow/scripts/optimizer.py
• LLMProvider (class) — Abstraction layer for different LLM API providers metagpt/provider/base_llm.py
• Config (class) — Centralized configuration management for the framework metagpt/config2.py
• CollectBlock (class) — Minecraft bot plugin for automated block collection and mining metagpt/environment/minecraft/mineflayer/mineflayer-collectblock/src/CollectBlock.ts

Sub-Modules

Configuration

config/config2.example.yaml (yaml)

• llm.api_type (string, unknown) — default: openai
• llm.base_url (string, unknown) — default: YOUR_BASE_URL
• llm.api_key (string, unknown) — default: YOUR_API_KEY
• llm.model (string, unknown) — default: gpt-4-turbo
• llm.proxy (string, unknown) — default: YOUR_PROXY
• llm.pricing_plan (string, unknown)
• embedding.api_type (string, unknown)
• embedding.base_url (string, unknown)
• +45 more parameters

config/config2.yaml (yaml)

• llm.api_type (string, unknown) — default: openai
• llm.model (string, unknown) — default: gpt-4-turbo
• llm.base_url (string, unknown) — default: https://api.openai.com/v1
• llm.api_key (string, unknown) — default: YOUR_API_KEY

config/examples/anthropic-claude-3-5-sonnet.yaml (yaml)

• llm.api_type (string, unknown) — default: claude
• llm.base_url (string, unknown) — default: https://api.anthropic.com
• llm.api_key (string, unknown) — default: YOUR_API_KEY
• llm.model (string, unknown) — default: claude-3-5-sonnet-20240620

config/examples/aws-bedrock.yaml (yaml)

• llm.api_type (string, unknown) — default: bedrock
• llm.access_key (string, unknown) — default: YOUR_API_KEY
• llm.secret_key (string, unknown) — default: YOUR_API_SECRET
• llm.region_name (string, unknown) — default: us-east-1
• llm.model (string, unknown) — default: meta.llama2-70b-chat-v1

Related Ml Inference Repositories

Frequently Asked Questions

Analyzed on March 31, 2026 by CodeSea. Written by Karolina Sarna.