If you read one line: Go is now a first-class language for building agents — Google's ADK for Go and Microsoft's Agent Framework for Go both ship first-party, with MCP, tool-calling, and multi-agent orchestration — but the two labs whose models you probably call most, OpenAI and Anthropic, still offer no official Go agent SDK. Choose your framework by who's committed to maintaining it, not by the language on the badge.

For two years, "build an AI agent" quietly meant "build it in Python, or maybe TypeScript." Every reference framework, every cookbook, every SDK sample assumed it. Go teams — the people running the concurrent, long-lived backends that agents actually need — wrapped a Python sidecar, shelled out to a subprocess, or reimplemented the loop by hand against a raw HTTP client. That era is ending, but unevenly, and the unevenness is the story.

What actually shipped#

In July 2026, Microsoft put its Agent Framework for Go into public preview, bringing the same Agent Framework concepts its .NET and Python developers already use to Go (Microsoft for Go Developers). The Go SDK covers the core: agents, tools, middleware, workflows, observability, and interoperability, with models from Azure OpenAI, Anthropic, and Gemini, plus tool-calling and MCP.

That lands Microsoft alongside Google, which brought its Agent Development Kit to Go — a model-agnostic, code-first framework optimized for Gemini but wired to OpenAI-compatible endpoints, Anthropic, Ollama, and Azure AI Foundry, with the A2A protocol and MCP Toolbox for 30+ databases (Google Developers Blog). ADK now spans four languages — Python, Go, Java, and TypeScript — and you pull it with a single go get google.golang.org/adk.

As The New Stack put it, Microsoft has now joined Google in backing Go for agents — while OpenAI and Anthropic lag.

The two catches founders should price in#

Catch one: the Go builds trail their siblings. Microsoft is explicit that declarative agents, RAG, and CodeAct aren't in the Go preview yet, and that .NET currently has broader product integrations. You're getting the runtime core — the loop, the tools, the orchestration — not the full feature surface that ships to Python and .NET first. If your roadmap depends on the newest declarative-agent or built-in-RAG conveniences, you'll be reimplementing them or waiting.

Catch two — the louder one — is who isn't here. The first-party push is coming from the framework and cloud vendors, not the model vendors. OpenAI's Agents SDK is Python and TypeScript only. Anthropic's official SDKs are Python and TypeScript, with Go left to a community client. So the two labs whose models the largest share of founders actually build on give Go developers no first-party agent library. You can still call GPT and Claude from Go — through ADK's provider adapters or a plain HTTP request — but you're doing it without the vendor's own batteries-included harness.

Why Go, and why now#

This isn't language fashion. An agent backend is a specific runtime shape: long-lived, heavily concurrent, I/O-bound, with many model calls and tool invocations in flight at once. That's the workload Go was built for — goroutines instead of a Python worker pool, a single static binary instead of a container full of interpreter and wheels, and a memory footprint that doesn't balloon under concurrency. Google saw that first; Microsoft followed. The frontier labs, whose center of gravity is research code, haven't — yet.

If you're weighing this against the community Go options, note the difference in maintenance risk: Eino (CloudWeGo) and LangChainGo are OSS you adopt on your own support terms. What changed in July is that two hyperscalers now put first-party engineering behind Go, which is a different bet to stake a product on. We compared the community layer in Eino vs LangChainGo, and the framework landscape more broadly in the agent-stack roundup.

The call#

If you run a Go backend and you've been paying the Python-sidecar tax, stop — pick the first-party framework whose backer matches your cloud and your model. Azure or .NET heritage points to Microsoft's Agent Framework for Go; a want for the most complete first-party kit points to Google's ADK. Either way, keep your OpenAI or Anthropic calls behind a thin provider interface, because that's the layer where Go is still second-class and where you'll want to swap in a first-party SDK the day one finally ships.