AgentKit by Inngest

AgentKit's unit of orchestration is the network. A network holds a list of agents, a shared state object and a router. On each loop iteration the router is called with the network, its state, the call stack and the last result; it returns the next agent to run or undefined to stop. If no router is given, a default LLM-based routing agent chooses. The docs recommend starting with code-based routers so that the order of agents follows explicit rules.

State has two parts: the message history of the run and a typed data object that tools and lifecycle hooks can read and write. The docs describe this as deterministic state-based routing, where agents fill in fields such as a plan or a done flag and the router reacts to them. State lasts only for one run(); persistence across runs comes from a history adapter that loads and saves threads, and long-term memory is shown through an integration with Mem0.

AgentKit can run as a plain library, but durable behaviour comes from Inngest. When a network runs inside an Inngest function, tool calls become retriable steps, LLM requests can be offloaded, tools can wait for external events, and runs appear with step-level traces in the Inngest dev server or cloud.

Human input is implemented as a tool. Inside the tool handler, Inngest's step.waitForEvent() pauses the network until an event with a matching field (for example a ticket ID) arrives or a timeout such as four hours passes; the event's data then becomes the tool result. This requires the network to run inside an Inngest function, since the handler receives step only in that context. The docs show a support network in which a technical-support agent waits for a developer's reply.

For chat UIs, the useAgent React hook exposes approveToolCall() and denyToolCall() for tool calls in an awaiting-approval state, plus cancel() to stop the current run. Code-based routers also let the application decide when a run should stop.

npm install @inngest/agent-kit inngest

import { createAgent, createNetwork, createState, createTool, openai } from "@inngest/agent-kit";
import { z } from "zod";

interface QAState { answer?: string }

const saveAnswer = createTool({
  name: "save_answer",
  description: "Store the final answer",
  parameters: z.object({ answer: z.string() }),
  handler: async ({ answer }, { network }) => { network.state.data.answer = answer; return "saved"; },
});
const researcher = createAgent({ name: "researcher", system: "Answer, then call save_answer.", tools: [saveAnswer] });
const network = createNetwork<QAState>({
  name: "qa", agents: [researcher], defaultModel: openai({ model: "gpt-4o-mini" }),
  router: ({ network }) => (network.state.data.answer ? undefined : researcher),
});

const state = createState<QAState>({});
await network.run("What does an AgentKit router do?", { state });
console.log(state.data.answer);

• inngest is a required peer dependency (>=3.43.1 in package.json); the installation page says so from v0.8.0 and the README from v0.9.0. zod must be version 4.
• Model adapters read provider keys from the environment, for example ANTHROPIC_API_KEY in the quick start.
• Local runs with tracing use the Inngest dev server (npx inngest-cli@latest dev).
• Human-in-the-loop tools only work when the network runs inside an Inngest function, because step is undefined otherwise.
• Typed state lasts for a single run(); pass a history adapter or your own state to continue a conversation.
• v0.4.0 moved anthropic() max_tokens into defaultParameters, a breaking change for older code.

Yes, as a client. Agents list MCP servers in mcpServers and use their tools over Streamable HTTP, SSE or WebSocket. No MCP server mode is documented.

inngest is a required peer dependency. Networks can run as a library, but retries, durable steps, traces and human-in-the-loop waits come from running inside Inngest functions.

The library is Apache-2.0. Inngest, the vendor, sells Inngest Cloud plans with a free tier and paid tiers for higher execution volumes.

The latest release, 0.13.2, came out on 2025-11-13, and the last push to the repository was on 2026-04-29.

A router is called after each agent. It can be a function that reads the network's typed state, or the default LLM routing agent; returning undefined stops the run.