Agentic AI frameworks

LangGraph comes from LangChain and models an agent as a graph of steps whose state is saved as it runs, so a run can resume after a failure or a pause. LangChain names J.P. Morgan, Uber and Klarna among the companies that use it. The OpenAI Agents SDK is the production successor of Swarm, an earlier experiment of OpenAI, and keeps to a few building blocks: agents with instructions and tools, handoffs between agents, guardrails that check inputs and outputs, and built-in tracing.

Microsoft Agent Framework reached version 1.0 for .NET and Python on 3 April 2026, with stable APIs and a promise of long-term support. Microsoft Agent Framework is the successor of Semantic Kernel and AutoGen, and Microsoft asks users of both projects to migrate. Google publishes the Agent Development Kit (ADK) for Python, TypeScript, Go, Java and Kotlin and states that it works with almost any generative AI model. Anthropic offers the Claude Agent SDK for Python and TypeScript, which gives an application the agent loop and the built-in tools of Claude Code, and a Claude agent workshop in Frankfurt starts its tooling with it.

The Model Context Protocol (MCP) is the common interface between an agent and a tool or a data source, and LangGraph, the OpenAI Agents SDK, Microsoft Agent Framework, ADK and the Claude Agent SDK can all call MCP servers. A bank can therefore write one server for a system and reuse it with any of these frameworks. The MCP servers of data providers and payment firms show which tools such a server exposes and what a bank checks before it connects one.

Agent2Agent (A2A) covers the link between agents: Google created the protocol, and the Linux Foundation took it over as an open project in June 2025, with SAP and ServiceNow among the founding contributors. Under A2A, two agents work together without exposing their internal memory, tools or logic to each other, so a bank agent can hand a task to the agent of a custodian or a payment provider. The Agent Payments Protocol extends A2A for payments that an agent starts on behalf of a person.

For high-risk AI systems, the EU AI Act requires automatic logs over the lifetime of the system and a person who can intervene or stop it through a stop button or a similar procedure. Deployers keep these logs for at least six months, and banks and other financial institutions keep them as part of the documentation that financial services law already requires. ESMA reminded investment firms that their management body stays responsible for decisions taken by people or by AI tools, and it expects records of how AI is used in investment services.

LangGraph pauses a run with an interrupt, for example before a financial transaction, saves the exact state through a checkpointer and continues once a person approves or rejects the step. Microsoft Agent Framework offers checkpointing, human approvals and pause and resume in all its orchestration patterns. The OpenAI Agents SDK records model calls, tool calls, handoffs and guardrail checks in traces that go to the OpenAI backend by default, so a bank that keeps its records in house replaces that trace processor with its own.

Most agent frameworks are open-source libraries that run inside the bank, and DORA treats such software as an information and ICT asset. In January 2026, BaFin published non-binding guidance on ICT risks in the use of AI that spells out what DORA means for such code. Software libraries belong in the inventory of ICT assets, libraries and frameworks are scanned for vulnerabilities on a regular schedule, and code from open-source projects is analyzed and tested before it reaches production.

BaFin names two risks of open-source libraries: malicious code that enters through a library, and a library whose maintainers stop after a few years while known vulnerabilities stay open. A release with stable APIs and long-term support, as Microsoft promises for Agent Framework 1.0, lowers the second risk. Hosted add-ons for tracing or deployment are ICT services from a third party and go into the register of information like any other provider.

Agentic AI frameworks with a stable public release include LangGraph from LangChain, the OpenAI Agents SDK, Microsoft Agent Framework, the Agent Development Kit from Google, the Claude Agent SDK from Anthropic and CrewAI. The first five can call tools through the Model Context Protocol, and Microsoft Agent Framework and ADK also support Agent2Agent, ADK still as an experimental feature.

In LangGraph the developer defines every step and branch as a graph, and a checkpointer saves the state along the way. The OpenAI Agents SDK has very few abstractions and leaves more of the routing to the model through handoffs. Microsoft Agent Framework combines free-running agents with graph-based workflows and runs on .NET as well as Python.

LangGraph, the OpenAI Agents SDK, Microsoft Agent Framework and CrewAI are published under the MIT license, and ADK and the A2A specification under Apache 2.0. The Claude Agent SDK falls under the commercial terms of Anthropic, and an open license never covers the language model behind an agent, which a bank licenses separately.

MCP connects one agent to tools and data, such as a market data feed or a payment API, while A2A connects agents with each other across companies and vendors. The A2A project calls the two complementary: a bank agent can use MCP for its own systems and A2A to pass a task to an agent outside the bank.

Python is the common language of all of them, and LangGraph, the OpenAI Agents SDK, ADK and the Claude Agent SDK also come in TypeScript. Microsoft Agent Framework is available for .NET and Python, with a Go version in public preview, and ADK adds Go, Java and Kotlin.