Google's Genkit Ships Agents API with Detached Turns and Human-in-the-Loop for TypeScript and Go

The core design of the Genkit Agents API is a single abstraction that spans different interaction modes. The same agent object can handle one-shot replies, drive streaming multi-turn conversations, pause for user approval of tool calls, or even manage detached long-running tasks. Developers no longer need to switch between different framework components for chatbots vs. complex multi-agent workflows.

The interface consolidates message history, tool execution loops, streaming, state persistence, and front-end protocols into a single chat() function. Whether the agent runs in-process or behind an HTTP endpoint, its behavior remains consistent. This design reduces the cost of migrating from prototype to production.

Genkit categorizes agent data into two types: custom state is typed application data driving the next turn (e.g., workflow state, task list); artifacts are generated outputs that can be independently reviewed, downloaded, or versioned (e.g., reports, code patches). Tools can update both during a session, with Genkit streaming changes to the client in real-time.

For persistence, the framework offers two paths. With a configured session store, the agent is server-managed: messages, state, and artifacts are persisted as snapshots, and clients reconnect using a session ID. Genkit includes built-in stores for Firestore (production multi-instance), memory (development), and file (local testing), with support for custom implementations. Without a configured store, the client-managed approach is used: the server returns the full state, and the client sends it back on each interaction. EBENEZER DON (AI engineer) notes that the client-managed mode is suitable for temporary sessions or strict data residency requirements, but at the cost of increased network payload as the conversation grows.

Detached turns allow a client to start an agent task and disconnect, then poll for results later. The agent continues working server-side, writing progress to snapshots that any client can read. For example, developers can create long-running research jobs or multi-step plans without needing a persistent WebSocket or a separate task queue. Example code shows tasks initiated via the detach() method, then polling snapshot state at 1-second intervals until completion.

Interruptible tools provide human-in-the-loop control with built-in anti-forgery protection. When a tool is marked as interruptible, the agent pauses mid-execution, returns a pending action to the client, and only continues after the user approves or rejects it. The runtime validates recovery requests against session history to prevent forged inputs. A Go example shows a security confirmation flow before shell command execution: if the command is risky, the tool interrupts and waits for user confirmation; if the user rejects, an error is raised.

For multi-agent coordination, Genkit’s middleware system (released in May) injects a delegation tool into each child agent, allowing the orchestrator model to route parts of a request to specialized sub-agents. Child agents can use the same chat() interface locally or be accessed via HTTP endpoints. The middleware layer also provides composable hooks, supporting exponential backoff retries, cross-provider model fallback, tool approval gating, and system prompts loaded via SKILL.md files.

Genkit remains model-agnostic through a plugin architecture. Official plugins support Google AI (Gemini), Vertex AI, Anthropic, OpenAI, and Ollama. A Vercel AI SDK adapter allows developers to integrate Genkit agents in Next.js apps. Each agent is itself a servable action, with route helpers providing turn, snapshot, and abort endpoints on standard HTTP endpoints.