Google Drops ADK 2.0 and Genkit Agents for AI App Builders

In a coordinated 24-hour release window, Google published three developer-facing launches: a new graph-based workflow runtime for the Agent Development Kit (ADK 2.0), a preview release of the Genkit Agents API for building full-stack conversational apps, and a Google Cloud Workbench extension for VS Code. Together, they represent Google's most concentrated push yet to own the agent-development stack end to end.

ADK 2.0: Deterministic Execution for Production Agents

The centerpiece is ADK 2.0, which ships a Workflow Runtime that transitions ADK from a hierarchical agent executor to a graph-based execution engine. The core argument is that production agents fail not because models are bad, but because teams reach for LLM orchestration when deterministic code would be cheaper, faster, and safer.

In benchmark results using Gemini 3.5 Flash and mock API responses, the ADK 2.0 workflow approach consumed roughly 50% fewer tokens per run (2,265 vs. 5,152) and ran 20% faster (5.7 vs. 7.2 seconds) compared to a vanilla autonomous agent handling the same task. The new release introduces a spectrum-of-control model: deterministic workflow nodes handle routing, tool calls, and structured handoffs, while LLM agents are reserved only for the nodes that require genuine reasoning — parsing unstructured input, drafting responses, making judgment calls.

"Developers shouldn't have to choose between flexibility and predictability," the ADK team wrote in their launch post. "They need the best of both."

ADK 2.0 Python has been generally available since May 19, 2026; ADK Go 2.0 reached general availability on June 30. Both support dynamic workflows, multi-agent collaboration, and human-in-the-loop gates built directly into the graph. Prompt injection resistance is a highlighted production benefit — because workflow edges are hardcoded rather than inferred by the model, a manipulated agent node cannot trigger unauthorized execution paths.

Genkit Agents: Full-Stack Conversational AI in Preview

Alongside ADK 2.0, Google previewed the Genkit Agents API for TypeScript and Go — an open-source, application-layer framework for building conversational agents inside existing web or mobile apps. The Agents API handles message history, state persistence, streaming, and human-in-the-loop interrupts behind a single chat() interface.

Developers define an agent on the server, pair it with a session store (Firestore for production, in-memory for tests), and drive it from any client through a remoteAgent() handle that speaks the same wire protocol. Detachable long-running turns let clients close a tab and reconnect by snapshot ID when the agent finishes — making multi-step research jobs and planning workflows practical without holding a connection open.

Google has drawn a clear boundary between the two frameworks: Genkit is for adding agentic features to an existing app; ADK is for building complex, standalone multi-agent systems on managed infrastructure. The two are explicitly composable — Genkit ships an adapter for the Vercel AI SDK's useChat hook, letting developers build their frontend from Vercel AI Elements while running Genkit on the backend.

Google Cloud Workbench VS Code Extension

A third July 1 release brought Google Cloud Workbench tooling into VS Code, giving developers access to Google Cloud's managed ML environment without leaving their editor. The extension is aimed at teams using Vertex AI or Gemini APIs who want tighter local-to-cloud integration during model development and iteration.

What's Unconfirmed

Google has not published independent token benchmarks for the ADK 2.0 workflow comparisons beyond its own illustrative data. The Genkit Agents API carries a preview label that explicitly warns of potential breaking changes in minor releases. The VS Code Workbench extension has not yet disclosed regional availability or pricing tier requirements.

For developers building agent-backed applications, this release slate is the clearest signal yet that Google intends to compete for the full developer toolchain — from the model and framework layer through runtime, IDE integration, and managed hosting.