Genkit's new Agents API simplifies building conversational full-stack AI apps

A framework for conversational AI at scale#

Genkit, Google's open-source framework for building AI-powered applications, is expanding its capabilities with a new Agents API that abstracts away much of the complexity required to create conversational, full-stack AI systems. The framework supports TypeScript, Go, Dart, and Python, and the new API is available in preview for TypeScript and Go.

Building sophisticated AI features—such as a customer support assistant that tracks context across multiple interactions, or a coding copilot that refines solutions over several turns—has traditionally required developers to manually implement message history, tool calling loops, streaming, state persistence, and client-server communication protocols. The Agents API consolidates this boilerplate into a single interface.

How the Agents API works#

The API centers on a simple concept: define an agent on the server with a name and system prompt, then drive it using an identical chat() interface whether it runs locally or remotely over HTTP. The agent can handle one-shot replies, streamed responses, paused tool calls, and multi-turn conversations without requiring developers to switch abstractions as features evolve.

Sessions and state management are configurable:

• Server-managed: The server persists messages, custom application data, and generated artifacts (like reports or plans) as snapshots. Clients resume by sending a session ID, making this approach ideal for persistent chat apps and shared devices.
• Client-managed: The server returns the full state; the client sends it back on the next turn. This suits applications that already handle persistence or require stateless server deployments.

Server-managed agents also enable conversation branching—users can explore alternatives from any saved point in history without disrupting the original thread.

State and artifacts#

Agents carry three categories of information across turns. Message history provides continuity; custom state holds compact, typed application data like workflow status or selected items that drive the next interaction; and artifacts are generated outputs users can inspect, download, or version separately, such as an itinerary or approved plan.