Decision comparison
AutoGen vs MetaGPT
AutoGen and MetaGPT both coordinate several agents on one task, with different degrees of specialisation. AutoGen is a general Python framework where agents converse to resolve a task, now in community-managed maintenance mode. MetaGPT assigns software-company roles — product manager, architect, engineer — that follow standard operating procedures to produce requirements, designs, code and tests.
AutoGen has a named successor from its own vendor
Microsoft merged AutoGen and Semantic Kernel into the Microsoft Agent Framework. AutoGen is in maintenance mode: it receives no new features or enhancements and is community managed going forward. Microsoft directs new projects to the Agent Framework.
Architecture choice. These take different approaches to the same problem. Read the table as a fit question rather than a feature race.
Applies to: Building multi-agent systems where agents take roles and collaborate.
All 2 are agent frameworks.
Quick Comparison
| Decision factor | AutoGen | MetaGPT |
|---|---|---|
| What it is | Microsoft's multi-agent framework, where agents hold a conversation to solve a task, now in maintenance mode | A multi-agent framework that assigns software-company roles — product manager, architect, engineer — following standard operating procedures |
| Model of work | Agents converse until a task is resolved, with group chat for more than two | Agents take software-company roles and follow standard operating procedures to produce artefacts |
| Output | Whatever the conversation produces, including code executed in a sandbox | Structured software artefacts: requirements, design documents, code and tests |
| Generality | A general multi-agent framework applicable to many task types | Specialised towards software development workflows |
| Vendor status | Community-managed maintenance mode; Microsoft's new work is in the Agent Framework | An independently developed open-source project |
| Control | You define agents, their system prompts and the conversation pattern | Roles and procedures come with the framework, with customisation on top |
| Best fit | Teams wanting general conversational multi-agent patterns in Python | Teams experimenting with automated software development pipelines |
AutoGen
- What it is:
- Microsoft's multi-agent framework, where agents hold a conversation to solve a task, now in maintenance mode
- Model of work:
- Agents converse until a task is resolved, with group chat for more than two
- Output:
- Whatever the conversation produces, including code executed in a sandbox
- Generality:
- A general multi-agent framework applicable to many task types
- Vendor status:
- Community-managed maintenance mode; Microsoft's new work is in the Agent Framework
- Control:
- You define agents, their system prompts and the conversation pattern
- Best fit:
- Teams wanting general conversational multi-agent patterns in Python
MetaGPT
- What it is:
- A multi-agent framework that assigns software-company roles — product manager, architect, engineer — following standard operating procedures
- Model of work:
- Agents take software-company roles and follow standard operating procedures to produce artefacts
- Output:
- Structured software artefacts: requirements, design documents, code and tests
- Generality:
- Specialised towards software development workflows
- Vendor status:
- An independently developed open-source project
- Control:
- Roles and procedures come with the framework, with customisation on top
- Best fit:
- Teams experimenting with automated software development pipelines
Public signals
Verified factual signals only. Bars appear only for like-for-like metrics with five weekly assessments for every tool; missing evidence stays explicit. These signals do not establish enterprise adoption, product quality, or total cost.
| Metric | AutoGen | MetaGPT |
|---|---|---|
| GitHub commits, 90d(Product adoption) | 0 | 0 |
| GitHub stars(Product adoption) | 60,000+ | 70,000+ |
| Search interest(Market interest) | 1 | 0 |
| Hacker News mentions, 90d(Community interest) | 3 | 0 |
| PyPI weekly downloads(Product adoption) | 87.4k | 1.2k |
| Stack Overflow questions(Community interest) | 37 | Not available |
As of September 14, 2026 — updated weekly.
Health & risk evidence
Observed public-source checks for mapped package versions and repositories.
AutoGen
September 14, 2026Package vulnerabilities
PyPI · autogen-agentchat@0.7.5
0 vulnerabilities
across 1 package
Repository security score
Not available
MetaGPT
September 14, 2026Package vulnerabilities
PyPI · metagpt@0.8.2
8 vulnerabilities
across 1 package
Repository security score
Not available
Feature Comparison
| Feature | AutoGen | MetaGPT |
|---|---|---|
| Coordination | ||
| Multiple agents on one task | Full support | Full support |
| Agent-to-agent conversation | Full support | Full support |
| Predefined role templates | Partial support | Full support |
| Human-in-the-loop steps | Full support | Full support |
| Output | ||
| Code generation | Full support | Full support |
| Sandboxed code execution | Full support | Full support |
| Document and specification generation | Partial support | Full support |
| Structured output enforcement | Full support | Partial support |
| Building | ||
| Python library | Full support | Full support |
| Tool calling | Full support | Full support |
| Model-agnostic across providers | Full support | Full support |
| Customisable agent roles | Full support | Full support |
| Status | ||
| Receiving new features | Not verified | Full support |
| Security and bug fixes | Partial support | Full support |
| Named vendor successor | Full support | Not verified |
| Large community | Full support | Partial support |
Coordination
Multiple agents on one task
Agent-to-agent conversation
Predefined role templates
Human-in-the-loop steps
Output
Code generation
Sandboxed code execution
Document and specification generation
Structured output enforcement
Building
Python library
Tool calling
Model-agnostic across providers
Customisable agent roles
Status
Receiving new features
Security and bug fixes
Named vendor successor
Large community
Which approach fits
AutoGen and MetaGPT both coordinate several agents on one task, with different degrees of specialisation. AutoGen is a general Python framework where agents converse to resolve a task, now in community-managed maintenance mode. MetaGPT assigns software-company roles — product manager, architect, engineer — that follow standard operating procedures to produce requirements, designs, code and tests.
When each approach fits
Choose AutoGen if:
Choose AutoGen when the problem is not specifically software development and you want to define the agents and the conversation pattern yourself. It is general, the patterns are well documented, sandboxed code execution is available when a task needs it, and the API is stable because development has stopped.
Choose MetaGPT if:
Choose MetaGPT when you are experimenting with automated software development and want role structure provided rather than designed. Predefined roles following standard operating procedures produce structured artefacts — requirements, design documents, code and tests — rather than a transcript you then have to interpret.
These scenarios reflect the available product evidence. Your requirements, existing stack, and team expertise should guide the final decision.
Frequently Asked Questions
What does maintenance mode mean for us?
Microsoft merged AutoGen and Semantic Kernel into the Microsoft Agent Framework, which reached 1.0 in April 2026. AutoGen is community-managed and receives no new features; Semantic Kernel v1.x continues to get critical bug and security fixes while new development goes to the Agent Framework. Both remain installable and both run in production today. The question that follows is how long you expect your code to live, and how much of it you would rewrite if you migrated.
Does role specialisation produce better results?
It produces more structured results, which is not the same thing. Predefined roles with procedures mean the output arrives as identifiable artefacts rather than a conversation to read, and that structure is genuinely useful. Whether the content is correct still depends on the model and the prompts. Review generated code as you would review code from an unfamiliar contractor.
Is automated software development ready for production use?
For well-specified, self-contained tasks with clear acceptance criteria, these systems produce usable starting points. For work that requires understanding an existing codebase, its conventions and the reasons behind past decisions, output needs substantial revision. Treat the result as a draft to review rather than as finished work, and measure how much editing it actually needs before scaling the approach.
How do we control cost on multi-agent systems?
Multi-agent designs multiply model calls: each agent reads context and produces output, and a conversation between four agents over ten turns is 40 calls before tools are counted. Set a per-run token budget, cap conversation turns, and use a small model for routing and a large one only where reasoning quality changes the answer. Without tracing, a prompt change that triples cost is invisible until the invoice.
What do these need to run?
Python in most cases, an API key for a model provider, and somewhere to run the process — a container on Kubernetes, a serverless function, or a service behind a REST endpoint returning JSON. Retrieval adds a vector store, often PostgreSQL with pgvector. Agents that execute generated code need a sandbox, usually a Docker container with no network access and a CPU limit, which is the one genuinely unusual infrastructure requirement in this category.
What does running one of these in production require?
Tracing every step with inputs, outputs, latency and token cost; a hard ceiling on iterations so a loop that fails to converge stops rather than spends; retries that distinguish a transient API error from a model producing unusable output; and evaluation that catches quality regressions when a prompt or model changes. Frameworks provide different amounts of that, and the gap is where the engineering effort goes.