ADK Go 1.0

Google's Agent Development Kit for Go hit 1.0 on March 31, 2026. It's a framework for building multi-agent systems in Go, with a focus on the operational side — tracing, security, extensibility — rather than just making it easy to call an LLM.

The headline features:

OpenTelemetry tracing is built in. You can visualize an agent's chain of thought alongside normal application metrics in Cloud Trace or any OTel-compatible backend. When something breaks, you can see whether the problem was a tool failure, a hallucination, or plain API latency. This is the kind of thing that separates "demo agent" from "production agent."

A plugin system for cross-cutting concerns. Instead of wrapping every tool call with logging and error handling, you inject plugins that intercept execution. The most interesting one is Retry and Reflect — when a tool call fails, instead of just retrying with the same parameters, the model gets the error, reconsiders, and adjusts. Self-correction without custom code.

human-in-the-loop confirmations. Tools can be flagged as RequireConfirmation, which pauses the agent and waits for a human to approve before executing. Intended for sensitive operations — financial transactions, database mutations, anything where an autonomous agent shouldn't just go ahead.

YAML configuration for defining agents. You can declare an agent's persona, model, sub-agent hierarchy, and tool access in YAML instead of Go code. Useful for iterating on agent behavior without recompiling.

agent2agent-protocol support for cross-language agent coordination — Go agents talking to Python or Java agents with structured message passing.

The repo is at github.com/google/adk-go.