Google ADK Review: Workflows, Pricing, and Practical Limits

This Google ADK review examines whether Google's Agent Development Kit fits an application that needs tools, controlled workflows, and a path to deployment. The short answer: it deserves a shortlist when you want a developer framework with substantial Google integration, provided you evaluate the exact language, runtime, and model combination you will operate.

ADK is an open-source framework, rather than a finished assistant or a subscription that includes inference. This documentation-based assessment was checked on November 2, 2026. It separates documented capabilities from practical recommendations and contains no firsthand speed, reliability, or cost benchmarks.

Google ADK Review: What ADK 2.0 Changes

An ADK agent combines a model, instructions, and optional tools. You can start with one agent and expand into workflows as routing, context, or code organization becomes harder to manage. That progression is more useful than beginning with a large team of agents simply because the toolkit supports one.

ADK 2.0 adds a graph-based workflow runtime. Google's current guide lists 2.0 availability for Python, TypeScript, and Go. Broader ADK language support should not be read as a promise that every new runtime feature exists in every implementation.

If recoverable execution is your main selection criterion, SD's LangGraph review offers a useful comparison. Evaluate both against the same interruption and approval scenarios, rather than comparing screenshots or counting orchestration features.

Existing ADK applications also need migration attention. Google's 2.0 guide identifies event-schema and custom-storage changes, among other compatibility considerations. A project that stores events in rigid database columns may require schema updates; strict downstream validators may need changes too. Treat an upgrade as an application change with recovery testing.

Workflows: Where Control Becomes Useful

Separate Model Reasoning from Ordinary Code

ADK's graph workflows connect agents, functions, and tools through explicit routes. The documentation distinguishes static graphs, programmatic dynamic workflows, and prebuilt workflow patterns. These approaches provide different ways to express a process; they do not establish that the resulting answers are correct.

Consider a purchase-request assistant. A model could extract the request from a message, an ordinary function could validate the budget code, and another step could prepare an approval summary. The budget check should remain deterministic when a conventional rule already expresses the requirement accurately.

Explicit routing makes the intended sequence easier to inspect. It also gives developers concrete boundaries for rejecting missing fields, stopping expensive branches, and recording why a run ended. Those are design opportunities, not automatic properties of every ADK application.

Plan the Human Decision

Google documents human-input nodes for graph workflows. This is relevant when a process needs information or an approval before it can continue. Your application still needs to decide who may respond and exactly which proposed action the response authorizes.

For the purchase example, show the amount, supplier, and requested action together. If any material detail changes afterward, request a fresh decision. An approval of an earlier summary should not become general permission for whatever a later model response proposes.

Model Support and Development Tools

ADK is optimized for Gemini, while Google's documentation also describes other model connections. Its Python LiteLLM integration can connect to remote providers or locally hosted models. That flexibility matters for an existing model contract or an application with local inference requirements.

Portability needs testing at the capability level. Check tool arguments, structured responses, streaming, and error handling with the actual adapter and model. A shared interface does not make every provider behave identically, and changing providers can change the workflow's failure modes.

Dependency maintenance is part of that choice. Google's LiteLLM connector page carries a security advisory about compromised historical package versions. Review the current advisory and dependency requirements before installation; copying an old tutorial's pinned environment can preserve problems that the current documentation already addresses.

ADK Web offers browser-based debugging, event history, and session-state inspection. Google explicitly limits it to development and debugging. Its visual builder is documented as Python-only. A working local interface therefore does not establish that you have a production application with appropriate access controls and operations support.

Before connecting sensitive systems, use SD's AI agent security guide to define tool permissions and review boundaries. Keep those rules enforceable in the application, where a model-generated instruction cannot silently relax them.

Google ADK Pricing: Framework and Service Costs

ADK's Python repository uses the Apache 2.0 license. There is no framework subscription charge for using that code, but a deployed application can incur model, compute, storage, and integration costs. Google Cloud services are separate from the framework license.

As checked on November 2, Google's published standard Agent Compute rate is $0.085 per vCPU-hour, with Agent Memory at $0.009 per GiB-hour. These are resource rates, not an all-inclusive ADK price. Google's pricing page also describes separately metered session and Memory Bank storage and operations.

Use the applicable service terms, region, discounts, and resource configuration when estimating your own bill. Model tokens remain another component. Evaluation runs and retries should appear in the estimate, rather than being treated as free background work.

A useful comparison metric is total cost per accepted task. Record all model calls and tool charges, then divide by tasks that meet a predefined acceptance rule. This is an evaluation method, not a measured ADK result. It helps expose designs that generate inexpensive individual responses but require frequent retries or human repair.

Deployment, Recovery, and Evaluation

Google documents deployment to Agent Runtime on Agent Platform, Cloud Run, GKE, and other container-friendly infrastructure. You can package an application for an environment without Google Cloud connectivity. However, its model endpoint and tools must also support that environment; container portability does not make remote dependencies disappear.

Choose deployment based on operational responsibilities. Identify who manages credentials, persistence, scaling, logs, and incident recovery. A managed runtime can change that division of work, while a container gives you more responsibility for assembling the surrounding service.

ADK's resume documentation describes configurable recovery for supported implementations and warns that custom agents need explicit support. Test the exact configuration you plan to ship. In particular, interrupt execution after a tool succeeds and before the workflow records completion. The surrounding application should prevent replay from creating a duplicate external action.

Evaluation is another substantial part of the toolkit. Google's evaluation guide, marked for Python, covers both final responses and execution trajectories. This distinction is valuable: an answer can look correct even when an agent consulted an unauthorized tool or took an unnecessary route.

Build a small evaluation set from normal requests, missing data, unavailable tools, and deliberately misleading retrieved content. Check permitted actions alongside output quality, latency, and cost. Keep that set stable when comparing models or upgrading the framework so that changes are visible rather than explained away by different examples.

Google ADK Alternatives and Who Should Shortlist It

ADK is a sensible candidate for developers who want Google's integrations and a toolkit spanning agent construction, inspection, evaluation, and deployment. It is less immediately useful for someone seeking a finished business assistant without application engineering.

For Python applications whose central concern is validated data and tool contracts, SD's PydanticAI review provides another comparison. Shortlist by the shape of your application, then check the precise recovery, approval, and deployment requirements on each framework's current documentation.

Run one representative workflow before committing to a broader platform decision. Use the same model where practical, the same tool permissions, and the same acceptance criteria. Record any feature that needs custom infrastructure. That exercise is more informative than an unsupported claim that one framework is universally easier or more production-ready.

Conclusion

This Google ADK review finds a credible framework to evaluate when controlled workflows and Google's ecosystem fit your project. The decision should rest on a representative task: verify language support, enforce tool boundaries, test recovery, and calculate the complete operating cost. Choose ADK when those results satisfy your requirements, then expand from the smallest working design.