Google’s Agent Development Kit (ADK) is an open-source, code-first framework for building, debugging, and deploying agents, including multi-agent systems. The Managed Agents API on Agent Platform is a different kind of tool: a REST-first way to create autonomous agents that run in isolated managed sandboxes. The deciding factor for an enterprise build is status. Google labels the Managed Agents API Pre-GA (before general availability), and under its stated terms it may not be used for commercial or production purposes. For production planning, ADK is the framework to build around, subject to checking its release and deployment-target status. The Managed Agents API is useful for evaluating agent behavior in a sandbox.
One correction to the title: Google’s current ADK overview describes the framework and its language support, but it does not name a “2.0” release. This article covers ADK as it is currently documented and does not attribute version-specific features or migration steps to it.
As an Amazon Associate I earn from qualifying purchases.
What Google ADK is
Google Cloud’s ADK documentation describes the framework in one sentence: “Agent Development Kit (ADK) is an open-source agent development framework that lets you build, debug, and deploy reliable AI agents at enterprise scale.” Read “at enterprise scale” as Google’s description of the framework’s intended use, not as a measured result.
Recommended Free Tools
ADK is code-first. You write the agent logic, define its tools, and control how work is orchestrated. The overview names workflow orchestration, dynamic routing, and multi-agent collaboration as supported patterns, along with tooling for evaluation. Google lists Python, TypeScript, Go, and Java as supported languages.
#1 Best Overall
On the deployment side, the overview lists three targets: Agent Runtime, Cloud Run, and Google Kubernetes Engine. Each has its own setup and operating model, so choose a target after the orchestration design is settled.
What the Managed Agents API is
The Managed Agents API overview introduces the product this way: “Managed Agents API on Agent Platform lets you build managed, autonomous agents with a single API call.” Read that sentence alongside the status limits in the next section. The offering runs an agent harness inside an isolated sandbox. Skills, files, packages, and connectivity are set as part of the environment configuration.
Two interfaces: Agents API and Interactions API
Google divides the product into two interfaces with different jobs.
- Agents API manages agent configurations and execution environments. This is the control plane, where you define what the agent is and what it can reach.
- Interactions API communicates with deployed agents at runtime. This is the data plane, where requests and responses flow.
When configuration changes, the system applies it to the sandbox. Source mounts and network allowlists are configurable, and those settings determine what the agent can reach.
Status and usage limits
Google’s Managed Agents API overview and lifecycle guide, reviewed on 7 October 2026 with the most recent update shown as 6 October 2026, classify the product as Pre-GA for limited testing and evaluation. Commercial and production use is not permitted under the stated preview terms. The guide also cautions against putting proprietary, sensitive, or confidential data into the service.
In practice, a prototype that exercises an agent’s tools and multi-step behavior fits the preview. A system that handles customer records, employee data, or confidential business material does not, even if the prototype works well. Preview stages and usage terms change, so recheck the Managed Agents API overview before planning any test that depends on them.
ADK and Managed Agents API compared
| Decision axis | Google ADK | Managed Agents API (Pre-GA) |
|---|---|---|
| Development model | Open-source, code-first framework; you write and control orchestration and tools (ADK overview) | Config-driven, REST-first creation of agents and their environments (Managed Agents API overview) |
| Orchestration | Workflow orchestration, dynamic routing, and multi-agent collaboration (ADK overview) | Autonomous agent harness in an isolated sandbox; multi-agent orchestration is not described in the overview |
| Where agents run | Agent Runtime, Cloud Run, or Google Kubernetes Engine (ADK overview) | Isolated managed sandbox on Agent Platform; no external network, system, or credential access unless configured |
| Languages | Python, TypeScript, Go, Java (ADK overview) | Not stated in the overview; the API is REST-first |
| Status | Described by Google as a framework for enterprise-scale agents; confirm the specific release and each deployment target at implementation time | Pre-GA; testing and evaluation only (see status section above) |
| Sensitive data | Not stated in the ADK overview; apply your own data-handling policy | Avoid proprietary, sensitive, or confidential data (see status section above) |
| Governance | Agent Platform identity, Registry, Gateway, Cloud Observability, and evaluation are available to configure | The same platform capabilities are relevant, but preview limits and sandbox access settings decide what is safe to test |
Which approach fits your situation
- Agents that customers, staff, or regulated data will touch in production: plan on ADK, and verify the release and deployment-target status before design sign-off.
- Observing how an autonomous agent behaves in an isolated sandbox with synthetic or non-sensitive data: the Managed Agents API preview fits this testing role.
- Custom orchestration, routing, and tool logic written in code: ADK is built for this.
- An undecided team: keep tool definitions, prompts, and evaluation cases in your own repository so the work carries over to either approach.
How the Managed Agents sandbox handles access
Managed Agents API agents have no access to external systems, networks, or credentials by default. Enabling an external API or an MCP tool requires an explicit developer configuration, and that connectivity adds risk you should control and monitor. Google’s guidance for doing this carefully is:
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →- Scope network reach narrowly, using network allowlists rather than open access.
- Use least-privilege access and short-lived credentials wherever the tool supports them.
- Monitor the actions the agent takes.
- Test each tool against sample or synthetic data first.
- Review critical outputs before anything downstream relies on them.
This guidance applies to controlled testing. It does not change the preview status described above.
Enterprise controls on Agent Platform
Google’s Agent Platform agents overview describes the governance layer around agents. Each control is a capability you configure. None is applied automatically to every agent.
Agent Identity
Each agent receives a unique Agent Identity in SPIFFE format, which you can reference in IAM policies to grant it specific permissions.
Agent Registry
Agent Registry centralizes metadata for agents and MCP tools, giving one place to see what exists.
Free tools Windows power users keep installed
One-click scans. No signup required.
Agent Gateway
Agent Gateway enforces policy on agent traffic.
Observability and evaluation
Cloud Observability provides traces, logs, and metrics for agent activity, and Gen AI evaluation supports testing agent quality. Set up both before a pilot reaches users, because incidents that were not captured cannot be traced afterward.
The same overview cites access to over 200 foundation models in Model Garden (Google Cloud, 2026). That figure is a catalog count, not a measure of model quality or enterprise adoption.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Securing hosted ADK agents registered with Gemini Enterprise
If you host an ADK agent on Agent Runtime and register it with Gemini Enterprise, the registration guide sets a specific rule for Model Armor: it must be configured through the REST API, inside the agent’s own application code. Model Armor settings in the Gemini Enterprise console do not automatically protect these ADK agents. Add this configuration to your release checklist, because it is easy to assume the console setting covers the agent.
Quick Recap
Enterprise readiness checklist for ADK
- Confirm the specific ADK release and the status of each deployment target in current Google Cloud documentation before design sign-off.
- Give each agent its own Agent Identity, and grant IAM roles scoped to the tools that agent calls.
- Record each agent and MCP tool in Agent Registry so ownership and inventory are visible.
- Apply Agent Gateway policy to agent traffic, and keep network reach limited to the allowlisted destinations you need.
- Enable Cloud Observability traces, logs, and metrics before the first pilot, and confirm they capture each tool call.
- Build an evaluation set from representative, approved data, and rerun it after every change to prompts, tools, or models.
- Configure Model Armor in the application code of every ADK agent registered with Gemini Enterprise.
What the evidence does not establish
- Adoption, performance, or cost figures. The pages reviewed publish none, so this article makes no claims about speed, accuracy, or pricing.
- Regional availability. The documentation reviewed does not cover availability by region, so confirm it for your project’s location.
- Production case studies. None are cited for either product.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
