Agent framework
A library or SDK that supplies the agent loop, state, memory, multi-agent patterns, approvals and tracing, so teams do not build them from scratch.
An agent framework is a library or SDK that provides the building blocks of AI agents: the loop that calls the model and runs tools, state and memory, coordination between several agents, human approval steps, and tracing.
What they have in common. Most frameworks cover the same ground under different names:
- Agent definition: a model, instructions and tools.
- Orchestration: graphs, crews, workflow agents or handoffs between agents.
- State and memory: per-conversation state and long-term stores.
- Human-in-the-loop: pause for approval or input, then resume.
- Observability and evaluation: traces, often in OpenTelemetry format, and test sets.
- Interoperability: MCP clients for tools, and increasingly A2A for working with agents in other systems.
Examples, checked September 26, 2026. Latest releases are from each project’s GitHub releases page. A2A support is as stated in each project’s own documentation.
| Framework | Maintainer | Latest release | A2A support in its docs |
|---|---|---|---|
| LangGraph | LangChain | 1.2.12 (September 21, 2026) | A2A endpoint in LangSmith’s Agent Server, v1.0 JSON-RPC binding |
| CrewAI | CrewAI | 1.15.22 (September 16, 2026) | Client and server modes; server cards default to protocol version 0.3.0 |
| Google ADK | Python 2.10.0 (September 25, 2026) | Expose and consume A2A agents, labelled experimental | |
| Microsoft Agent Framework | Microsoft | Python 1.19.0 and .NET 1.22.0 (September 18, 2026) | Host agents over A2A and call remote A2A agents |
AutoGen, Microsoft’s earlier multi-agent framework, is in maintenance mode, and its README directs new users to Microsoft Agent Framework.
Choosing one. Frameworks differ most in how much structure they impose: explicit graphs of nodes and edges, role-based crews with a manager, or model-driven delegation between agents. The choice does not bind the other side of an exchange. An agent built with one framework can work with an agent built with another over A2A, as long as both speak a compatible protocol version.
Neighbouring terms. LangGraph, CrewAI, Google ADK and AutoGen each have their own entry.
Sources
- LangGraph releases on GitHub (accessed )
- LangSmith documentation: A2A endpoint in Agent Server (accessed )
- CrewAI releases on GitHub (accessed )
- CrewAI documentation: Agent-to-Agent (A2A) Protocol (accessed )
- Agent Development Kit (Python) releases on GitHub (accessed )
- Agent Development Kit documentation: ADK with Agent2Agent (A2A) Protocol (accessed )
- Microsoft Agent Framework releases on GitHub (accessed )
- Microsoft Learn: Host agents with A2A (Microsoft Agent Framework) (accessed )
- AutoGen README on GitHub (accessed )