WritingGuide

Google ADK integration: evaluate, integrate, operate

For teams that already have systems, data and a process to automate, and want to know whether Google ADK fits and what it asks of them after launch.

I'm Matteo Gazzurelli, fractional CTO and agentic AI consultant in Brescia, Google Partner through MaGa Srl (opens in a new tab). This is the checklist I use to evaluate ADK, connect it to a client's systems and operate it after launch.

What Google ADK is, in three lines

Google ADK (Agent Development Kit) is an open-source framework for agents, tools and workflows, available in Python, TypeScript, Go, Java and Kotlin. For a map of ADK and Gemini Enterprise Agent Platform, read Google ADK in 2026.

When does Google ADK make sense?

I say yes to ADK when the four conditions below hold. When one is missing, the first conversation is about that condition, before any code.

Your team writes code

ADK puts agent logic in your repository, with tests and reviews. Your team needs developers who can maintain the code and operate the integrations.

Your systems are reachable

The agent needs an API, a database or an MCP server to act on. A process that lives in email threads and spreadsheets gets a data step first, then an agent.

Google Cloud is a reasonable home

ADK agents run anywhere a container runs; on Google Cloud they deploy to Agent Runtime, Cloud Run or GKE. If your data and identity already sit there, the scaffolding is in place; otherwise that integration is the first cost to weigh.

The work needs judgement across steps

A fixed sequence of calls is a workflow. ADK earns its place when the next step depends on what the previous one found.

Three cases where I advise against ADK: a one-off task, a process that still changes every month, a team already standardised on another framework that serves it well.

How I integrate ADK with your systems

  1. Tools before agentsI wrap the APIs you already have as ADK tools, and connect systems that expose an MCP server through ADK's MCP support. The agent sees only the actions I give it.
  2. One agent per responsibilityA single agent with a clear job, sub-agents only when the job splits cleanly. When agents owned by different teams have to talk, ADK speaks the A2A protocol (still labelled experimental in ADK).
  3. A person on the costly stepsEvery action that writes to a system of record, sends a message or spends money goes through an approval step, in the instructions and in code.
  4. Evals before launchA set of recorded cases the agent must pass, run with ADK's evaluation tooling (the adk eval command, the Eval tab of the ADK web UI, or pytest; Python only today) before the first deploy and again on every change.
  5. Deploy where you already operateAgent Runtime on Gemini Enterprise Agent Platform when you want the managed service; Cloud Run or GKE when your operations team already runs containers there.

What running it asks of your team

Traces you can read

Each run leaves a trace of model calls, tool calls and agent steps: ADK emits OpenTelemetry traces and can send them to Cloud Trace (documented for Python, TypeScript and Go); someone on your team reads the failed ones every week.

Evals as regression tests

The eval set from step four runs on every change to a prompt, a tool or the model version. A change goes live when it passes.

Model versions pinned

The Gemini model is pinned to a version and upgraded on purpose. ADK also runs other models through LiteLLM, so a model change is a configuration change with an eval run attached.

Your team owns it

Before I step back, your developers have deployed a change end to end: code, eval, deploy, trace. The handover is complete when that has happened once.

Google or something else?

I also build with Claude Agent SDK. Claude Managed Agents is a separate managed service. I compare the development and hosting options in Google vs Anthropic for AI agents, using the same review date as this guide.

Related

Integrating ADK into something real?

Write me a few lines on the systems involved and the process you want to automate. A reply within 2 business days.

Request an intro