AgentIndex icon
AgentIndex

Open-Source Multi-Agent Frameworks: CrewAI, Pydantic AI, Google ADK, Microsoft Agent Framework & More

Picking an agent framework in 2026 is less about prompt templates and more about software engineering: how tools are typed, how state is checkpointed across long runs, and whether you want autonomous role delegation or a deterministic graph.

One shift matters in particular: Microsoft has placed AutoGen in maintenance mode and directs new projects to Microsoft Agent Framework. Meanwhile, code-first toolkits like Pydantic AI, CrewAI, Google ADK, and AgentScope 2.0 have matured around native tool calling, MCP, and A2A.

AgentIndex · Updated 2026-10-20

Multi-agent frameworks by design style

Choose by whether you want role teams, typed Python loops, or enterprise cloud SDKs.

  1. Role teams & Flows

    Persona delegation and SOP pipelines.

  2. Typed & composable Python

    Pydantic validation, ReAct blocks and A2A.

  3. Cloud & multi-language SDKs

    Google ADK and Microsoft Agent Framework.

Quick picks

  • CrewAI — combines autonomous role-based agent teams (Crews) and deterministic event-driven workflows (Flows) in Python (MIT).
  • Pydantic AI — type-safe Python agent loop from the Pydantic team with dependency injection, validated outputs, and Temporal support.
  • adk-python — Google's Apache-2.0 code-first toolkit with graph workflows, multi-agent hierarchies, a dev UI, and evaluation tooling.
  • agent-framework — Microsoft's active successor to AutoGen for Python and .NET, with checkpointing, OpenTelemetry, and DevUI.
  • agentscope — Apache-2.0 Python 3.11+ toolkit with transparent ReAct blocks, A2A, realtime voice, and strong support for Chinese and global LLM APIs.
  • multica — issue-driven multi-agent workspace in TypeScript/Docker that coordinates and reviews up to 26 coding agent CLIs.

Side-by-side comparison

ToolWhat it isStarsLicenseLast pushMaintenanceReleases, 90 daysCommits, 12 weeksInstall
CrewAIPython framework for role-based multi-agent workflows60.0kMIT2026-10-21Active24Commits, 12 weeks: 32, 20, 22, 26, 28, 33, 37, 12, 16, 27, 34, 23uv tool install crewai
Pydantic AIType-safe Python agent framework from the Pydantic team20.8kMIT2026-10-22Active30n/auv add pydantic-ai
adk-pythonGoogle's code-first Python toolkit for building AI agents21.9kApache-2.02026-10-22Active18Commits, 12 weeks: 109, 137, 100, 131, 108, 89, 96, 132, 130, 138, 159, 26pip install google-adk
agent-frameworkMicrosoft's Python and .NET framework for agents and workflows14.3kMIT2026-10-21Active26Commits, 12 weeks: 33, 32, 70, 37, 74, 121, 132, 86, 116, 82, 118, 1pip install agent-framework
agentscopePython framework for transparent, production-ready agents33.6kApache-2.02026-10-21Active7Commits, 12 weeks: 16, 11, 32, 17, 26, 33, 29, 30, 46, 49, 22, 0uv pip install agentscope
MetaGPTMulti-agent framework that simulates a software company71.0kMIT2026-01-21Slowing0Commits, 12 weeks: 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0pip install --upgrade metagpt
Microsoft AutoGenMicrosoft's multi-agent framework, now in maintenance mode61.5kMIT2026-10-16Active0Commits, 12 weeks: 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0pip install -U "autogen-agentchat" "autogen-ext[openai]"
deer-flowByteDance's open-source super-agent runtime for long tasks84.3kMIT2026-10-22Active1Commits, 12 weeks: 32, 42, 27, 61, 80, 91, 134, 187, 190, 259, 176, 72uv sync
multicaAssign issues to AI coding agents like teammates53.5kn/a2026-10-22Active30Commits, 12 weeks: 164, 143, 183, 105, 96, 112, 107, 111, 45, 21, 46, 30curl -fsSL https://raw.githubusercontent.com/multica-ai/multica/main/scripts/…
rufloAgent meta-harness for Claude Code and Codex swarms78.2kMIT2026-10-21Active30Commits, 12 weeks: 0, 63, 23, 3, 10, 55, 59, 126, 375, 630, 233, 99npx ruflo@latest init wizard
nanobotUltra-lightweight self-hosted personal AI agent framework49.3kMIT2026-10-21Active2Commits, 12 weeks: 104, 121, 148, 96, 68, 90, 89, 118, 96, 128, 107, 43curl -fsSL https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/insta…

Tools in this list with an in-depth AgentIndex review: 11/11

How to choose

  • Decide between role-based delegation (Crews/SOPs) and code-first typed loops or graphs.
  • Check language requirements: most frameworks are Python-only, whereas Microsoft Agent Framework gives first-class support to both Python and .NET.
  • Verify active maintenance: avoid starting new production systems on frameworks in maintenance mode (such as AutoGen) or repos whose commit activity has stalled.
  • Look at how state and durability are handled—for example, built-in workflow checkpointing or Temporal integration.

FAQ

Should I still start a new project on Microsoft AutoGen?

No. Microsoft has placed AutoGen in maintenance mode with no new feature development and recommends starting new projects on Microsoft Agent Framework, which includes an official migration guide.

When should I pick Pydantic AI over CrewAI?

Pick Pydantic AI when you want a clean, type-safe Python agent loop where tools and structured outputs are validated by Pydantic. Pick CrewAI when your problem naturally splits into distinct agent personas collaborating inside event-driven Flows.

Which frameworks support languages other than Python?

Microsoft Agent Framework offers first-class Python and .NET packages (plus a separate Go repository), and Google's Agent Development Kit (ADK) publishes editions across multiple languages.

How should engineering teams coordinate multiple autonomous coding agents?

Instead of running agents in separate local terminals without audit trails, use an issue-driven workspace like Multica. It assigns GitHub issues to agents, enforces human review gates before merging code, and tracks execution costs and tool calls per run.

How we compiled this list

Tools are chosen by hand from the projects in our index, favouring those we have reviewed in depth. Stars, licence, last-push date and install command come from each repository and are refreshed with the site data. The maintenance column comes from the GitHub API (archived flag, last commit, releases in the last 90 days) and is refreshed daily. We read each project's README and docs; we have not run every tool ourselves, so test before you depend on one.

More lists