Google’s Agent Development Kit, or ADK, provides software for building agent applications. It should be assessed separately from a model such as Gemini and from the service used to host your application.
This guide explains the approach through an illustrative task. It does not rank ADK against other frameworks or claim tested production results.
What ADK provides
The official documentation covers agents, tools, orchestration, sessions, state, evaluation and deployment. Its current documentation includes explicit graph-based workflows alongside model-directed work. Google ADK documentation.
These capabilities address different responsibilities. A model proposes steps. A workflow can determine sequence. Your tool service performs the requested operation. Hosting determines how the application runs and accesses resources.
Check support for your chosen language and release. A capability documented for one implementation should not be assumed to behave identically in another.
Design a mixed workflow
For a fictional support process, use fixed steps to validate a case and record the final outcome. Between those steps, let an agent investigate order and tracking information with permitted tools.
Keep the approval decision explicit. The investigation should produce a draft and evidence. Sending can remain a separate operation available only after the required decision.
This structure gives you a testable division: fixed code controls the business sequence; model-directed work handles the variable investigation. It does not establish that ADK is the best choice for your organization.
Decide how state is used
Record which case is active, which sources were checked and whether a draft exists. Distinguish those operational facts from text supplied to the next model call.
Decide which data survives a restart. Ensure one user’s session cannot supply another user’s case context. Treat retained memory as a separately governed data store.
If execution is interrupted after a write, check whether the destination received the operation before retrying. Framework state alone does not establish whether an external transaction completed.
Tool access needs application controls
Give the model narrow interfaces with clear required fields. An order lookup and a save-draft operation should have different permissions from sending a reply.
Determine the acting identity at the tool service. Do not derive authority from model-generated text. Apply business limits at the service that executes the action.
If you connect an external server, assess its operator, authentication, data handling and network destination. Standardized communication does not establish that the server is appropriate for your task.
Choose hosting deliberately
Document where the runtime, state and artifacts reside. Confirm how credentials are provided, who updates the application and who responds to failed jobs.
Avoid assuming a successful local prototype proves the hosted deployment behaves the same way. Repeat relevant access, interruption and network tests in the intended environment.
Include review time and failed attempts when measuring cost. A faster model response does not necessarily reduce the time required to complete the case.
Prototype checklist
- Draw fixed and model-directed steps separately.
- Choose a supported language and record the version.
- Define session ownership and persistent state.
- Enforce access in tools and destination services.
- Test hosted behavior and interrupted actions.
- Compare completed tasks using the same evaluation cases as other candidates.
Use the official documentation for exact APIs and deployment steps. This guide supplies the questions your prototype should answer.
Testing whether an agent completes the task ↗
Measure results, permissions, recovery and cost with repeatable normal and difficult cases.
Prepared with AI assistance and checked against the linked documentation. Examples and numerical limits are illustrative unless stated otherwise. These guides do not report independent product testing. Check current documentation before choosing a tool.