Agent Development Kit

ADK ships in Python, TypeScript, Go, Java, and Kotlin, and the same concepts carry across all five. A team writes agents in the language it already uses, which removes a rewrite before anyone commits. Python has the widest examples and docs, so it's the usual starting point.

You can build a system of specialized agents and let them hand off tasks to each other, either through model-driven routing or an explicit tool call. Workflow agents handle the predictable parts: run steps in order, in parallel, or in a loop until a condition is met. That mix lets a team keep deterministic logic where it matters and let the model decide where it doesn't.

New in ADK 2.0, graph workflows weave deterministic code with adaptive reasoning in a structured graph, so each execution path is explicit and outcomes are easier to predict. For agents that touch money or compliance, that predictability is the point. It's Google's answer to the complaint that agent behavior is hard to pin down.

ADK is tuned for Gemini but works with most generative models through its model interface, including models from other providers and ones running locally. Teams on Google Cloud get managed access to a wider model range with controls for cost and performance. You aren't locked to one vendor's model if your needs change.

The framework includes a command line tool and a Developer UI for running agents, stepping through events, inspecting state changes, and tracing what happened. You can build multi-turn evaluation datasets and score results locally. Debugging an agent that made a strange decision is the difference between days and hours here.

Agents run anywhere that takes a container, from a local Docker image to Cloud Run, GKE, or Google's managed Agent Runtime. Deploying to Google Cloud adds managed scaling, authentication, and tracing without rewriting the agent. Develop locally, scale globally.