LangGraph, CrewAI, and AutoGen are the three most-reached-for open-source frameworks for building multi-agent LLM applications. LangGraph models agents as nodes in a graph with explicit state, CrewAI organizes them into role-based crews, and AutoGen structures collaboration as conversation between agents. They differ most in control, learning curve, and state handling.
This comparison is for builders choosing among them — and for the growing group who will discover they need something different from all three. We will be straight about both.
LangGraph: graphs and explicit state
LangGraph, from the LangChain ecosystem, treats a multi-agent application as a state machine: agents and tools are nodes, edges define what runs next, and a typed state object threads through every step. Control flow is explicit — branches, retries, and human-in-the-loop pauses are all visible in the graph rather than buried in prompts.
That explicitness is the draw and the tax. Teams that need auditable, deterministic-ish flows — where you can point at the graph and say exactly what happens after a failed review — accept the learning curve. Teams prototyping an idea often find the ceremony heavy for a first version.
CrewAI: role-based crews
CrewAI starts from an organizational metaphor: define agents as roles with goals and backstories, group them into a crew, hand the crew tasks. The framework decides much of the execution flow for you, which makes it the fastest of the three from zero to a working multi-agent demo — the configuration reads like a team description, not a program.
The convention that makes it fast also bounds it. When you need precise control over ordering, retries, or state shape, you are working against the framework's defaults rather than with them. CrewAI fits well-understood, repeatable processes — content pipelines, research-summarize-report patterns — better than intricate engineering workflows.
AutoGen: agents as conversation
AutoGen, from Microsoft, models multi-agent work as a conversation: agents exchange messages, critique each other, and converge on an answer, with patterns like group chats and nested conversations as the building blocks. For open-ended problems — explore this dataset, debate this design — conversational dynamics surface perspectives that a rigid pipeline would never generate.
The flip side is predictability. Conversations meander, token costs scale with chatter, and reproducing a specific outcome is harder than in a graph. AutoGen rewards research-flavored work and experimentation more than production pipelines with strict SLAs.
How to choose
Choose LangGraph when control and auditability dominate — you are encoding a process you already understand and need it to run the same way twice. Choose CrewAI when speed to a working crew matters and your process fits the role-team metaphor. Choose AutoGen when the problem is open-ended and agent-to-agent dialogue is genuinely the mechanism you want. All three assume one more thing: that you are building the agents.
What none of them solve: agents you don't build
The multi-agent reality for most software teams in 2026 is not custom-built agent graphs — it is Claude Code in one terminal, Cursor in another editor, and Codex on a teammate's machine, all touching the same application. You do not control those agents' internals, so there is nothing for an in-process framework to orchestrate. The coordination problem moves outside the process: shared task state, shared contracts, and claims that every tool can see.
That layer is what AppHandoff provides — a hosted MCP server that IDE agents and framework-built agents alike connect to for scanned API contracts, mismatch detection, and a shared ticket board with role tracking. Framework agents can use it the same way coding agents do, so a LangGraph pipeline and a Claude Code session stop being separate worlds. The architecture behind that pattern is covered in MCP server architecture, and the day-to-day workflow in multi-agent orchestration across Claude Code, Cursor, and Codex.
Frameworks orchestrate the agents you build. An orchestration platform coordinates the agents you use. Most teams scaling AI development end up wanting both — and they compose cleanly.