Field note · Google Cloud agents · July 2026
Google ADK in one day: learn the framework, then engineer the production boundary
Google ADK makes a TypeScript agent and typed function tools quick to build. The differentiator in forward-deployed AI is the system around the agent: product judgement, authenticated tools, evaluation, observability, adoption and operational ownership.
Users, triggers, decisions, systems and risk.
The smallest release with measurable value.
Typed tools, identity, state and integration.
Evaluation, UAT, adoption and operations.
What ADK gives an engineer quickly
Google's TypeScript quickstart uses @google/adk, FunctionTool, LlmAgent and Zod-defined parameters. That gives TypeScript teams a code-first structure for agent instructions, tools, sessions and deployment rather than hiding application contracts inside prompt text.
My implementation lab uses three bounded tools to capture a current workflow, define the smallest useful release and draft acceptance criteria. Every result remains a reviewable draft. The agent has no authority to change a system of record.
The forward-deployed pattern starts before the model
The first questions are about who performs the work, what starts it, where judgement is required, which systems and permissions are involved, and how value will be measured. ADK can structure evidence and prepare options, but the engineer still owns product judgement and the business remains responsible for acceptable operational risk.
This distinction prevents a framework tutorial from being mistaken for a customer outcome. The useful deliverable is not an agent endpoint. It is an adopted workflow with a clear operating boundary.
Scope the smallest useful release
Forward-deployed delivery reduces a broad AI ambition to one workflow that users can validate. Define the user, trigger, decision, approved data, permitted actions, success measures and explicitly deferred scope. A narrow release produces evidence sooner and limits the number of unknowns introduced at once.
For the lab, the boundary is deliberately modest: discovery notes become a structured current state, a minimum release proposal and a draft acceptance plan. There is no autonomous production action.
Tools are the real application boundary
Each production tool needs a narrow purpose, least-privilege identity, schema validation, tenant and user context, logging, idempotency, confirmation for consequential actions and explicit failure behaviour. ADK documents OAuth 2.0 and OpenID Connect flows for authenticated tools, including interactive user consent when required.
ADK also supports MCP integration, including using MCP servers as tools and deploying agents with MCP connectivity. MCP standardises connectivity; enterprise identity, authorisation, evidence and recovery still belong around the tool plane.
Evaluate behaviour, not only prose
ADK's evaluation documentation emphasises the final answer and the trajectory used to reach it. Documented criteria include exact tool-trajectory matching, reference-response comparison, rubric-based response and tool-use quality, and groundedness checks.
The current evaluation guide is primarily documented through the Python ADK tooling, so a TypeScript delivery should verify SDK feature parity and use an external evaluation harness where needed. The production target remains the same: test tool selection, arguments, permissions, missing-data behaviour, escalation and business task completion.
- Deterministic schema and environment checks
- Representative workflow cases and expected tool trajectories
- Rubric-based quality and groundedness evaluation
- Adversarial permission and prompt-injection scenarios
- Latency, model cost, tool cost, retries and human overrides
Deployment is not adoption
Google documents TypeScript deployment to Cloud Run through npx adk deploy cloud_run. The service can require authenticated access, and the development UI is optional rather than the production experience.
Production readiness also needs secrets, network boundaries, monitoring, incident ownership, rollback and support procedures. Adoption requires evidence that users complete the workflow faster or more consistently, understand the agent's limits and know when to override or escalate.
What transfers from other agent platforms
The production principles transfer directly from LangGraph, Amazon Bedrock and AgentCore, Microsoft Foundry and other orchestration stacks: explicit state, typed tools, governed retrieval, human authority, evaluations tied to business outcomes, correlated traces and safe recovery.
Learning ADK is not about placing another framework name on a profile. It demonstrates that the same engineering discipline can be applied across model and cloud ecosystems.
Executive takeaway
A capable engineer can learn the basic ADK API quickly. The valuable skill is knowing what to build around it. In forward-deployed AI, the framework creates the agent; product judgement, security, evaluation, adoption and operational engineering create the outcome.
Implementation lab
The downloadable TypeScript lab demonstrates three structured tools, explicit sequencing and a human-review boundary. It is intentionally small enough to inspect and extend.
Workflow capture, minimum-release scoping and acceptance planning with typed Zod contracts.