Who this is for: teams moving an agent from prototype to production in a firm that will be asked to show what the agent did. Skip it if you are prototyping: almost any framework will do for a demo.
In short: Framework choice in 2026 is mostly a cloud and vendor decision, not a technical one. All the serious options speak the same tool protocol. What separates them for a regulated team is whether they can pause for a human, survive a crash without losing state, and run on more than one model.
Key points
- LangGraph ranks first for checkpointing that lets runs recover from crashes and pause for approval at any node.
- Microsoft Agent Framework 1.0 reached general availability on 3 April 2026, merging AutoGen and Semantic Kernel.
- Claude Agent SDK has the most complete MCP integration but runs Claude models only.
- OpenAI Agents SDK keeps conversation history in built-in sessions, but durable checkpointing of long workflows is largely yours to build.
- Before choosing, kill a run midway, test a logged human-approval step, and swap the model.
How we ranked
| Criterion | Weight | Why it matters |
|---|---|---|
| State and checkpointing | 30% | An agent that cannot resume after a failure, or pause at a named step for approval, cannot be supervised. |
| Human-approval gates | 25% | Supervisors now expect a review step that leaves evidence. |
| Model and cloud neutrality | 25% | Locking orchestration to one model vendor is the hardest dependency to unwind later. |
| Cost transparency | 20% | Multi-agent patterns multiply model calls quietly. |
The list
1. LangGraph
- Best for: teams that need durable, inspectable workflows across models.
- Why it ranks first: it models an agent as a state machine with checkpointing, so a run can recover from a crash and pause for human approval at any node. That maps directly onto what a supervisor will ask for.
- Watch out for: more design work up front than role-based frameworks, and the commercial tracing tools around it are a separate decision.
- What would change its rank: another framework matching its checkpointing with less engineering.
2. Microsoft Agent Framework
- Best for: firms standardised on Azure and Microsoft 365.
- Why it ranks here: version 1.0 reached general availability on 3 April 2026, merging AutoGen and Semantic Kernel into one supported product, with the enterprise identity and governance tooling Microsoft sells around it.
- Watch out for: cost of conversational multi-agent patterns. The comparison we drew on notes that a four-agent, five-round group chat is at least 20 model calls.
- What would change its rank: evidence of production use outside the Microsoft stack.
3. Claude Agent SDK
- Best for: teams building on Claude that want the deepest support for the shared tool protocol.
- Why it ranks here: the most complete integration with MCP, the connection standard major financial data providers adopted in 2025–26.
- Watch out for: it runs Claude models only, so it fails the neutrality test by design.
- What would change its rank: support for other model providers.
4. Google Agent Development Kit (ADK)
- Best for: teams on Google Cloud that need agents to talk to other agents.
- Why it ranks here: native support for A2A, the agent-to-agent protocol that reached a stable 1.0 with more than 150 supporting organisations in April 2026.
- Watch out for: a heavy dependency on Google Cloud for production deployment.
- What would change its rank: easier deployment outside Google Cloud.
5. OpenAI Agents SDK
- Best for: simple, handoff-based agents on OpenAI models.
- Why it ranks here: a clean handoff model that is quick to learn and widely documented.
- Watch out for: built-in sessions keep conversation history, but durable checkpointing and resumption of long workflows are largely yours to build.
- What would change its rank: native durable state.
Left out, and why
- CrewAI. Role-based and quick to prototype with, but criticised for extra model calls and less suited to supervised production work.
- Writing your own. One large financial-data firm built a custom orchestrator that runs JSON-defined workflows, which grew far beyond its first version, and later weighed moving to a managed service. Building your own is defensible at that scale. At most others it is a maintenance bill.
Three questions before you choose
- Kill the process halfway through a run. Does it resume from the last step, and is the state inspectable?
- Can a named step wait for a human, and is that approval logged with who, when and what they saw?
- Swap the model. How much code changes?
Common questions
What is the best AI agent framework for production in regulated industries?
LangGraph ranks first, because it models agents as state machines with checkpointing, so runs can recover from crashes and pause for human approval at any node.
What is Microsoft Agent Framework?
It merges AutoGen and Semantic Kernel into one supported product. Version 1.0 reached general availability on 3 April 2026 and suits firms standardised on Azure and Microsoft 365.
Does the OpenAI Agents SDK support state persistence?
Partly. Built-in sessions keep conversation history, but durable checkpointing and resumption of long workflows are largely yours to build.
How should I test an agent framework before choosing it?
Kill the process mid-run and check it resumes with inspectable state, check that a named step can wait for a logged human approval, and see how much code changes when you swap the model.
Disclosure: lowob has no commercial relationship with any framework on this list. Facts checked 17 Oct 2026.