Agent Org Chart of the Day · No. 04 · MetaGPT
Agents hand off documents, not chat.
Platform: MetaGPT (framework and paper) · Roles: 5 · Source: Hong et al., arXiv 2308.00352
Why it matters
MetaGPT runs a simulated software company as an assembly line with standard operating procedures. Every handoff is a structured document: a PRD, a system design, a task list, code and tests, published to a shared message pool that each role subscribes to by profile. The paper puts it plainly: agents communicate through documents and diagrams rather than dialogue. The Engineer closes the loop with executable feedback, running its own code against tests for up to 3 retries. The open question for any team: what is the handoff document, and is it agreed before the work starts?
Roles and reporting lines
| Role | Input | Output (handoff document) |
|---|---|---|
| User (human) | - | Requirement |
| Product Manager | User requirement | PRD: goals, user stories, competitive analysis, requirement pool |
| Architect | PRD | System design: file list, data structures, interface definitions, sequence flow |
| Project Manager | System design | Task list per code file |
| Engineer | Tasks and design | Code; writes tests, runs, debugs (max 3 retries) |
| QA Engineer | Code | Unit tests and review |
| Shared message pool (store) | All roles publish | Each role subscribes by profile |
Routines
- One sequential SOP run per project
- Publish to the shared message pool; each role subscribes to what its profile needs and acts once its prerequisite inputs arrive
Approval gates
| Where | Check | Documented |
|---|---|---|
| Engineer | Executable feedback: testing continues "until the test is passed or a maximum of 3 retries is reached" | Yes |
| QA Engineer | Tests and review before output | Yes |
| Human sign-off | - | Not described in the paper |
Copyable spec
org_chart:
name: MetaGPT software company (SOP assembly line)
number: 4
platform: MetaGPT (MIT); pattern portable to any stack
source: https://arxiv.org/abs/2308.00352
repo: https://github.com/FoundationAgents/MetaGPT
credit: Hong, Zhuge, Chen, Zheng, Cheng, ... Wu, Schmidhuber (DeepWisdom, KAUST and others)
roles:
- {id: pm, title: Product Manager, input: user requirement, output: "PRD (goals, user stories, competitive analysis, requirement pool)"}
- {id: arch, title: Architect, input: PRD, output: "system design: file list, data structures, interface definitions, sequence flow"}
- {id: pjm, title: Project Manager, input: system design, output: task list per code file}
- {id: eng, title: Engineer, input: tasks + design, output: code, loop: "write tests, run, debug; max 3 retries"}
- {id: qa, title: QA Engineer, input: code, output: unit tests + review}
communication:
style: structured documents, not dialogue
bus: shared message pool; each role subscribes to what its profile needs
trigger: an agent acts only after its prerequisite inputs arrive
approval_gates:
- {where: Engineer, check: executable feedback (tests pass, max 3 retries), documented: true}
- {where: QA, check: tests and review, documented: true}
- {where: human, documented: "no sign-off step described"}Sources
- Sirui Hong and 14 other authors, MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework, arXiv 2308.00352 (submitted 1 Aug 2023, v7 revised 1 Nov 2024). Fetched 15 Oct 2026, 22:14 CEST.
- FoundationAgents, MetaGPT repository. Fetched 15 Oct 2026, 22:14 CEST.
What is confirmed / what is not
Confirmed (paper v7):
- the five roles; the SOP workflow
- communication "through documents and diagrams (structured outputs) rather than dialogue"
- shared message pool + subscription mechanism
- executable feedback, "a maximum of 3 retries"
- repo: MIT license (GitHub)
Not confirmed:
- benchmark results used specific vendor models → no benchmark numbers quoted
- repo last push 21 Jan 2026 (GitHub API, 15 Oct 2026) → "framework and paper", not "actively shipping"
- no human approval step described; chart shows automated checks only