Choose a Framework by the Mechanism You Need

The OpenAI Agents SDK, LangGraph, Microsoft, and CrewAI offer different levels of control. Compare not slogans but state, topology, durability, approvals, and observability.

Tool · Documented focus · Check before choosing

• OpenAI Agents SDK - Agents, tools, handoffs, guardrails, sessions, HITL, and tracing; Provider fit, storage ownership, approval integration
• LangGraph - Graph state, persistence, subgraphs, workflows, and multi-agent patterns; State schema, reducers, deployment model
• AutoGen AgentChat - Agents and predefined teams: round robin, selector, swarm, Magentic-One; Termination, shared context growth, serialization
• CrewAI - Crews for collaboration and Flows for event-driven control/state; The Flow/Crew boundary, telemetry, operational ownership

Microsoft: AutoGen and Agent Framework. AutoGen AgentChat is documented (round robin, selector, swarm, Magentic-One, termination), but Microsoft merged AutoGen and Semantic Kernel into Agent Framework - the current unified product (GA 1.0 since April 2026 for .NET and Python, Go in preview) with graph-based Workflows, checkpointing, and human-in-the-loop. For a new project look at Agent Framework, and treat AutoGen material as its predecessor; before choosing, check the current SDK and the migration guide.

To choose a starting surface by the mechanism you need, use the lab.

The lab's result is a shortlist, not a ranking. Take a minimal use case and check the behavior under tracing: pause/resume, schema validation, cancellation, retries, parallel branches, budget accounting, and upgrade path.

A framework doesn't remove system design. It cuts boilerplate but doesn't know your trust boundaries, business acceptance, cost of error, and data ownership. Don't enable a topology just because it exists in one constructor.

The framework is chosen. What remains is not to confuse it with the protocols: orchestration is neither MCP nor A2A.