Google Genkit v.Agents: detached turns and human-in-the-loop are now in preview

In plain terms. Google has released a preview of the Genkit Agents API for TypeScript and Go. It wraps message history, tool loop, streaming, and persistence behind a single chat(). Two standout features are detached turns (the agent continues even after the client disconnects) and a HITL with anti-forgery validation on resume.

InfoQ (July 14, 2026) reports the preview release. Genkit already aimed at orchestrating prompts and flows; adding a stateful agent layer changes the game for teams that don't want to recode a custom harness. The interesting primitive is the detached turn: the API knows that an inference turn can last several minutes, that the client (browser, worker) will lose connection, and that a server-side job store must keep control. It's the missing piece between "instant chat" and "long-running agent that reminds the user in twenty minutes."

The HITL is not a marketing checkbox: interruptible tools require anti-forgery proof on resume, human validation, and the call resumes with a signed token. This is exactly what agent post-mortems in production have been asking for—a year to properly stop a tool call and resume without re-emitting the entire state.

Under the hood - The preview API exposes, according to the documentation quoted by InfoQ: unified TS/Go chat(); pluggable message history and state persistence; streaming; interruptible tool loop; detached turns via server-side job identifier. For a custom harness already using LangGraph or a custom queue, the real question is: is Genkit's chat() surface extensible enough to connect your memory store (pgvector, DynamoDB) and your policy engine? In preview, the answer is probably "yes for standard cases, no for the edge."

So what. Google is fueling the market for pre-built harnesses (see also Adaptive Recall, the MCP memory wave, the arXiv "Better Harnesses, Smaller Models"). For a CTO shipping an agent in 2026, the question shifts from "which model?" to "which harness?" Genkit joins the shortlist alongside AI SDK, LangGraph, and Semantic Kernel. The risk: Google Cloud lock-in on persistence and authentication. To watch: the stable availability of the API, and whether detached turns hold up under load for 5-30 minute agents in production.