Microsoft Agent Framework — The Successor to AutoGen

From AutoGen to Microsoft Agent Framework

Until recently, this unit would have been about Microsoft AutoGen -- the research framework that made "agents talking to each other" famous. That story has changed, and the change itself is a lesson in technology risk.

AutoGen is now in maintenance mode. Its own project page describes it as community managed, receiving bug and security fixes only, and points new users to its successor: Microsoft Agent Framework. Microsoft built Agent Framework by combining AutoGen's multi-agent orchestration with the enterprise plumbing of Semantic Kernel, Microsoft's other agent toolkit. It was previewed in October 2025 and reached version 1.0 -- generally available for .NET and Python, with stable APIs and long-term support -- in April 2026. Semantic Kernel 1.x continues to receive critical fixes for at least a year after that release.

The irony is worth noting for any vendor review: AutoGen, once Microsoft's flagship agent project, is now community-supported. The long-term support commitment sits with Agent Framework.

Five Built-In Ways for Agents to Work Together

AutoGen was known for one idea: agentsAgentsAI systems that can autonomously plan and execute multi-step tasks by calling tools, querying data sources, and making decisions without human intervention at each step -- typically within defined permissions and human approval checkpoints.See glossary as participants in a conversation, debating and refining until they reach a conclusion. Agent Framework keeps that idea but makes it one option among five built-in patterns:

• Sequential: agents work in a fixed order, each passing its output to the next -- like a document moving through a checklist
• Concurrent: several agents work on the same task in parallel and their results are combined
• Handoff: one agent passes the case to a better-placed specialist, the way a branch refers a client to the private bank
• Group chat: agents discuss a problem together, challenging and building on each other's contributions
• Magentic-One (manager-led): a manager agent plans the work, assigns tasks to other agents, and tracks progress

Agent Framework also supports graph-style "workflows" with conditional routing, similar to LangGraph, so teams are no longer forced to choose between a conversational style and a structured one.

The Group Chat Pattern

The group chat remains the most distinctive pattern for banking analysis. Multiple agents, each with its own role and instructions, take turns presenting their perspective, responding to the others, and refining the analysis until the conversation reaches a conclusion.

Closing the Predictability Gap

The main criticism of AutoGen-style conversations was that they could be hard to predict and hard to stop. Agent Framework adds the controls a bank would expect:

• Human approval: a workflow can pause until a person approves a step, such as sending a customer communication
• Checkpointing: the state of a long-running workflow is saved, so it can survive a restart and be inspected later
• Pause and resume: a workflow can wait -- for a document, a reviewer, or an overnight batch -- and pick up where it left off

A Note on Agents That Write Code

AutoGen popularised agents that write and run their own analysis code -- for example, querying a data warehouse and producing statistics for a portfolio question. Whatever framework you use, treat code-executing agents as a higher-risk category: they need an isolated sandbox, limited data access, and logs of every piece of code they ran.

The Microsoft Ecosystem, Without the Same Lock-In

Microsoft Foundry

The managed home for Microsoft's agent work is Microsoft Foundry (renamed from Azure AI Foundry in November 2025). Since July 2026, Foundry Agent Service's "hosted agents" can run Agent Framework or LangGraph code as a managed service -- so a bank can write agents in a framework of its choice and still have Microsoft operate them.

Multi-Provider by Design

Unlike AutoGen's early close ties to Azure OpenAI, Agent Framework works with many model providers: Foundry, Azure OpenAI, OpenAI, Anthropic's Claude, Amazon Bedrock, Google's Gemini, and locally hosted models through Ollama. It also supports MCP, the open standard for connecting agents to tools; support for A2A, the agent-to-agent standard, was announced as following the 1.0 release. That makes the "tied to Microsoft" trade-off weaker than it used to be.

Enterprise Support

For many banks the deciding factor is the vendor relationship: enterprise agreements, service commitments, and account teams that banking institutions expect for critical infrastructure. With Agent Framework, that relationship comes with a long-term support commitment -- something AutoGen no longer has.

Considerations for Banking

Strengths

• Several collaboration patterns in one framework, including group chat for multi-perspective deliberation
• Human approval, checkpointing, and pause-and-resume for controlled, auditable workflows
• Long-term support from Microsoft, with a natural home in Foundry for Microsoft-centric banks
• Works with many model providers, which reduces model-layer lock-in

Trade-offs

• Younger than LangChain, with a smaller talent pool and less mature retrieval (RAG) tooling
• Conversational patterns still need careful design and testing to keep agents on task
• A2A support was still arriving at the 1.0 release -- confirm the current status before relying on it
• Teams with AutoGen or Semantic Kernel code face a migration project