Article Talk Edit Source View History

Open Source Agents

A working list of the open source agent frameworks worth knowing, as of April 2026.

THIS LIST WILL AGE
This field moves quickly. Treat this article as a snapshot, not a permanent ruling. The Council of Builds reviews it quarterly. The last review concluded: "Still too many frameworks. Still the right ones on top."
CONTENTS
  1. Overview
  2. Tier 1: Production-Ready
  3. Tier 2: Worth Knowing
  4. Tier 3: Watch List
  5. Picking One
  6. See Also
  7. References

Overview

An open source agent framework is a library or toolkit for building LLM-powered systems that can plan, call tools, and execute multi-step workflows autonomously. Gartner predicts 40% of enterprise applications will feature task-specific AI agents by end of 2026, up from less than 5% in 2025.[1] That means the choice of framework matters; it also means there are, currently, far too many of them.

What follows is a curated short list. The full market has dozens of options; this is the subset Burbridge would actually pick from for a production build, grouped by readiness.

Tier 1: Production-Ready

Battle-tested, well-documented, broad community. Pick from this tier first.

FRAMEWORK BEST FOR NOTE
LangGraph Stateful multi-agent workflows Enterprise leader; ~34.5M monthly downloads. Cyclic-graph model for agents. The default pick when the workflow is complex.
OpenAI Agents SDK Single-vendor builds, clean ergonomics Polished. Opinionated. Model-locked to OpenAI unless you wrap it — which you should, see model-agnostic architecture.
AutoGen (Microsoft) Conversational multi-agent orchestration Models agent collaboration as a dynamic conversation between agents. Good when the problem is genuinely multi-voice.
CrewAI Fast prototyping of multi-agent systems Role-based. Shared goals. Fastest path from idea to running multi-agent system — developers report 2–4 hour prototypes.
Google ADK Prototype-to-production on Google Cloud Agent Development Kit. Built-in multi-agent orchestration, tool use, session management. Good if you're already in Google's ecosystem.

Tier 2: Worth Knowing

Strong in specific niches. Pick from here when the use case matches.

FRAMEWORK WHY
Dify Leads in GitHub stars (~130k). Low-code platform; good for teams with designers and non-engineers who need to contribute.
Mastra TypeScript-first. If your stack is Node/TS end-to-end, this removes the Python detour.
Langflow Open-source low-code for RAG + multi-agent. Visual builder. Good for prototyping with stakeholders in the room.
Smolagents (Hugging Face) Minimalist. Agents write code to call tools rather than emitting JSON — clever, works well when the LLM can code.
Pydantic AI For people who like their types. Structured output is a first-class citizen. Excellent for back-office agent work.

Tier 3: Watch List

Worth tracking. Not yet where you'd stake a business-critical system, but moving fast.

  • OpenHands (formerly OpenDevin) — open-source coding agent; improving rapidly.
  • SWE-agent — academic (Princeton), software-engineering-focused, strong on benchmark performance.
  • MetaGPT — multi-role agent framework; interesting design ideas, less production uptake.
  • Goose (Block) — local-first agent, privacy-friendly, developer-focused.
  • Aider — AI pair programming in the terminal; narrower scope, but in that scope, excellent.

Picking One

A short decision tree that will get you 80% of the way there:

  • Complex stateful workflow, graph-shaped? → LangGraph
  • Conversational multi-agent, Microsoft stack? → AutoGen
  • Need a prototype today, role-based team? → CrewAI
  • Google Cloud shop, want production plumbing? → Google ADK
  • TypeScript / Node end-to-end? → Mastra
  • Low-code, non-engineers contributing? → Dify or Langflow
  • Minimalist, code-generating agents? → Smolagents
  • You want strict types and structured output? → Pydantic AI

Whatever you pick, apply model-agnostic architecture and treat the framework the same way: the framework is also pluggable. Today's right choice may not be right in 18 months. Build the seams where you can replace it.

"Pick the framework. Then build it so the framework is also replaceable. That is not paranoia. That is framing." — Burbridge

See Also

Model-Agnostic Architecture · Software · Aggressive Craftsmanship · The One Ring (on over-ambitious general-purpose tools) · Foundation

References

  1. Gartner prediction cited in Firecrawl's Best Open Source Frameworks For Building AI Agents in 2026.
  2. Lindy. (2026). Top 11 AI Agent Frameworks (2026): Expert-Tested & Reviewed.
  3. Shakudo. (2026). Top 9 AI Agent Frameworks as of March 2026.
  4. Langfuse. (2025–2026). Comparing Open-Source AI Agent Frameworks.
  5. Firecrawl. (2026). Best Open Source Frameworks For Building AI Agents in 2026.
CATEGORIES:People Who ShipBuilders