Microsoft Agent Framework for Go: What the Public Preview README Says Is Built, and What Is Not

Microsoft Agent Framework for Go is the Go implementation of Microsoft's agent framework, and its README says it is in public preview. The GitHub repository shows an MIT license and releases v0.1.1 (October 13, 2026), v0.2.0 (October 20) and v0.2.1 (October 27). If you write Go and were waiting for a first-party agent framework, the useful question is which pieces exist today. The README answers that fairly directly, including a list of things marked "not implemented yet". It also contradicts itself on one point, which I cover below.

I read the README, the .NET and Go feature comparison doc, and the release notes for v0.2.0 and v0.2.1, and checked repository metadata through the GitHub API. I did not install the module or run the examples. Everything below about behavior is what the documents say.

What it is, per the README

• An "open, multi-language framework for building production-grade AI agents and multi-agent workflows". The Go repository holds the Go implementation, samples and docs. The .NET and Python versions live in the separate upstream repository.
• A note says the Go version "is in public preview and is currently evolving outside the core upstream codebase", with closer alignment expected as adoption grows.
• Install is go get github.com/microsoft/agent-framework-go.
• It lists support for Microsoft Foundry, Azure OpenAI, OpenAI and Model Context Protocol, among others.

What the README lists as available

• Middleware for request and response processing, logging, OpenTelemetry, context providers, tool approval and automatic tool calling.
• Workflows: graph-based patterns including sequential, concurrent, group collaboration, conditional routing, subworkflows, checkpointing, streaming, human-in-the-loop and time-travel.
• Foundry agents: project-backed agents, invoking existing server-side agents, client headers, served-model metadata and Foundry memory.
• Observability through OpenTelemetry.
• Agent skills, text search, chat history memory, persistent chat history in Cosmos DB, embeddings and vector data.

What the README marks as not implemented

• Handoff orchestration.
• Foundry-hosted deployment.
• Declarative agents.
• AF Labs.
• DevUI.

The "Preview status" section adds: "Declarative agents, RAG, CodeAct, and functional workflows are not yet available."

Where the README disagrees with itself

The Key Features list includes "Text Search (RAG)": retrieve and cite external knowledge before agent runs or through an on-demand search tool. The Preview status section says RAG is not yet available. Both are in the same file. Either the status note is stale or "RAG" there means something narrower than the text search feature. I did not resolve it, and I would test the text search package before relying on it.

What v0.2.0 and v0.2.1 changed

The v0.2.0 release notes describe embeddings, vector search and memory:

• A provider-independent embedding package with ordered batch results, numeric and binary vectors, model metadata and token usage, plus OpenAI and Foundry generators that produce float32 embeddings, and opt-in OpenTelemetry instrumentation that does not capture raw inputs or vectors.
• A vectordata package with shared contracts for typed and dynamic collections, CRUD, portable filters and vector search, with in-memory, Azure AI Search and Azure Cosmos DB for NoSQL connectors.

The v0.2.1 notes list small changes, including decoding structured output after tool execution and preserving A2A task status messages in agent responses. Several entries are labeled as .NET porting work, some by a bot account. That fits the comparison doc's framing of the Go SDK as a port with a symbol catalog against .NET.

What the comparison doc is, and is not

The feature comparison doc describes a catalog that maps .NET symbols to Go counterparts, with statuses such as mapped, adapted, partial, unmapped and intentional. Its own warnings say the API extraction scope is incomplete and that "symbol correspondence does not establish behavioral parity". It also says extraction-added placeholders marked unreviewed are not assessed gaps. So the catalog is a tracking tool, not a statement that the Go SDK matches .NET. The README says in short that Go covers core agents, tools, middleware, workflows, observability and interoperability, while .NET has broader product integrations and several features not implemented in Go.

Data and responsibility notes in the README

• The repository "does not configure a Microsoft telemetry exporter". The OpenAI and Foundry adapters add SDK identity and version to the User-Agent, and add a feature-usage bitmask only on approved HTTPS Azure or Foundry destinations. The README says the mask records which capabilities were used, not counts, prompts, arguments or user data.
• A notice says that if you connect to third-party servers, agents, code or non-Azure models, you do so at your own risk and are responsible for the data that leaves your Azure compliance boundaries, and for your own responsible AI mitigations and testing.

What I would check

These are my untested suggestions.

1. List the features your project needs and compare them to the not-implemented list, especially handoff orchestration.

2. Resolve the RAG question by reading the text search package and running its example.

3. Run a workflow with checkpointing and kill it midway to see what restart does.

4. Check whether the in-memory vector store is enough for your tests before adding Azure services.

5. Look at the User-Agent feature mask on your network if you have egress rules.

6. Pin a version, since the module is at v0.2.x.

What I would not conclude

• That Go has parity with .NET. The README says .NET has broader integrations and features not implemented in Go.
• That a mapped symbol behaves the same. The comparison doc says it does not establish behavioral parity.
• That it is ready for production use. The README describes the goal as production-grade agents, but labels the Go version public preview.
• That RAG is or is not available. The README says both.

Frequently asked questions

What is Microsoft Agent Framework for Go?

The Go implementation of Microsoft Agent Framework, in public preview, with agents, tools, middleware, workflows, observability and memory.

What license does it use?

MIT, per the GitHub repository metadata.

What is not implemented yet?

Handoff orchestration, Foundry-hosted deployment, declarative agents, AF Labs and DevUI, per the README. The Preview status section also lists RAG, CodeAct and functional workflows.

What did v0.2.0 add?

Embeddings, vector data contracts and connectors for in-memory, Azure AI Search and Cosmos DB, per the release notes.

Does it send telemetry to Microsoft?

The README says the repository configures no Microsoft telemetry exporter. It describes a feature-usage bitmask added to the User-Agent on approved Azure and Foundry HTTPS destinations.

Can I use non-Azure models?

The README lists OpenAI and other providers. It also says use of third-party systems is at your own risk.

Did you run it?

No. I read the documentation and release notes.

Sources

All read on November 2, 2026.

• microsoft/agent-framework-go README
• .NET and Go SDK feature comparison
• Releases and repository metadata (GitHub API)

Not read: the Go source, the examples, the Transparency FAQ, the Microsoft Learn docs for the framework, and the .NET and Python repository.