Full-Code · Orchestration Frameworksactive

MetaGPT

Type: full-code · Vendor: DeepWisdom / FoundationAgents · Language: Python · License: MIT · Status: active · Status in practice: mature

Links: homepage docs repo

Materialise software-engineering Standard Operating Procedures as multi-agent teams of Roles whose communication is mediated by an Environment.

Description. MetaGPT is DeepWisdom's open-source multi-agent framework whose stated philosophy is 'Code = SOP(Team)' — it encodes the SOPs of a software company as a Team of Role agents (Product Manager, Architect, Project Manager, Engineer, Data Analyst). Roles observe Messages from a shared Environment and emit Actions; the Environment broadcasts messages to subscribed roles until everyone is idle. MetaGPT has also published the AFlow paper on automatic agentic workflow generation (ICLR 2025 oral) and the Data Interpreter agent that writes and runs code to solve analytics tasks.

Agent loop shape. Role-based multi-agent loop. A Team contains Role agents, each with a list of Actions and a watch list of Message types to subscribe to. The shared Environment broadcasts messages to subscribed roles; each role observes, decides which Action to run, executes it, and publishes a Message back. The Team runs for n_round steps or until env.is_idle (no role has new messages).

Primary use cases

code generation from a one-line requirement via a simulated software company
multi-role SOP-encoded multi-agent workflows
data analysis via the Data Interpreter code-as-action agent
automatic workflow search via AFlow

flowchart TD REQ[One-line requirement] --> TEAM[Team / company] TEAM --> ROLES{Roles PM / Architect / PrjM / Engineer} ROLES --> PM[Product Manager Role] ROLES --> ARCH[Architect Role] ROLES --> ENG[Engineer Role] ROLES --> DA[Data Interpreter code-as-action] PM -->|publish Message| ENV[(Environment broadcast bus)] ARCH --> ENV ENG --> ENV ENV -->|subscribed types| ARCH ENV -->|subscribed types| ENG ROLES -->|_act / Action| OUT[Artifacts: specs / code / docs] ENV -.env.is_idle.-> STOP[Stop] TEAM -.n_round limit.-> STOP TEAM -.AFlow.-> SEARCH[Automatic workflow search]

Key concepts

Code = SOP(Team) (docs) — Core philosophy: a Standard Operating Procedure for a software team is materialised as code executed by an LLM team.
Role → role-assignment (docs) — Agent abstraction with a system prompt, a list of Actions, and a watch list of Message types it subscribes to.
Environment → inter-agent-communication (docs) — Shared message bus that broadcasts published messages to roles based on their subscriptions; tracks env.is_idle for termination.
Team / company.hire / company.run → sop-encoded-multi-agent (docs) — Container that wires Roles together, invests budget, and runs the SOP for n_round rounds.
Data Interpreter → code-as-action (docs) — Code-as-action agent that writes and executes Python code to solve data analysis tasks.
AFlow → automatic-workflow-search (docs) — Automated agentic workflow generation (ICLR 2025 oral) shipped in MetaGPT.