Google ADK

What it is

Google's Agent Development Kit (ADK) is an open-source, code-first framework for building multi-agent systems, in Python and Java. Agents are objects defined in code rather than diagrams or config, tools are ordinary functions with schemas, and the whole thing is deliberately model-agnostic — it is built to work with Gemini, but the model is a swappable part rather than the point.

The word doing the work in the spec above is deployment-ready. Plenty of frameworks will get you to a working demo; ADK is aimed at the part after that: sessions with real state, evaluation harnesses, and supported paths onto Google's serving infrastructure. If you are already on Google Cloud, the distance from local prototype to hosted agent is noticeably shorter.

How it works

An agent has instructions, a model and a set of tools — function tools, built-ins, or tools reached over protocols such as MCP. Agents can hand work to other agents, and the kit ships workflow agents for the boring, important shapes: sequential, parallel and loop. Those are worth having, because “run these three in order, then repeat until clean” is what half of real orchestration actually is, and writing it as a state machine each time is waste.

Around the agents sits the operational layer: sessions that carry state across turns, artifacts for files and generated objects, callbacks for control at the edges of a call, and an evaluation facility that runs a set of examples against a candidate configuration. The result is a framework that reads more like application code than like a prompt playground.

When it earns its place over a plain loop

When the system has more than one agent, or when it needs to survive contact with users. The workflow primitives cover the orchestration cases that otherwise turn a plain loop into ad-hoc control flow, and the session and artifact handling answers the questions every production loop eventually hits: where does state live between turns, and where do generated files go? If the answer to “where does this run in six months?” includes Google's infrastructure, that tips it further.

For a single agent with a handful of tools, it is more framework than you need. A plain loop with typed function calls is simpler to read and to debug, and it has fewer moving parts to keep current.

Limits

• Young and moving. ADK is a relatively recent entry and its API has changed quickly; anything specific on this page should be checked against the docs before you build on it in late 2026.
• Two languages, not one twice. The Python and Java surfaces are both advertised; whether they are at parity at any given moment is exactly the sort of thing to verify rather than assume.
• Deployment gravity. The serving story is strongest on Google's own platform. Model-agnostic is not the same as deployment-agnostic.
• Abstraction tax. Sessions, artifacts and workflow agents are real machinery; if you do not need them, they are code you have to understand before you can change a prompt.

Alternatives on this board

• LangGraph — explicit graph control flow when the workflow shapes are not enough.
• OpenAI Agents SDK — a smaller multi-agent API with handoffs and guardrails.
• CrewAI — role-based crews when the split is organisational rather than procedural.
• Managed Agents — hosted agent infrastructure if you would rather not run it at all.