agent.ts - aixyz

The agent definition file. Must export a default ToolLoopAgent and optionally accepts for payment gating and capabilities for A2A agent card configuration.

app/agent.ts

import { openai } from "@ai-sdk/openai";
import { stepCountIs, ToolLoopAgent } from "ai";
import type { Accepts } from "aixyz/accepts";
import type { Capabilities } from "aixyz/app/plugins/a2a";
import weather from "./tools/weather";

export const accepts: Accepts = {
  scheme: "exact",
  price: "$0.005",
};

export const capabilities: Capabilities = {
  streaming: false,
  pushNotifications: false,
};

export default new ToolLoopAgent({
  model: openai("gpt-4o-mini"),
  instructions: "You are a helpful weather assistant.",
  tools: { weather },
  stopWhen: stepCountIs(10),
});

Exports

Export	Type	Required	Description
`default`	`ToolLoopAgent`	Yes	The agent instance
`accepts`	`Accepts`	No	Payment config — gates the A2A `/agent` route
`capabilities`	`Capabilities`	No	A2A capabilities — controls streaming and push notification support

When accepts is exported, the /agent endpoint requires x402 payment. Without it, the agent is not registered on the A2A endpoint. accepts can also be an array of payment entries for multi-network support.

Capabilities

The optional capabilities export controls the A2A agent card’s capabilities field and the executor’s behavior:

import type { Capabilities } from "aixyz/app/plugins/a2a";

export const capabilities: Capabilities = {
  streaming: false, // default: true
  pushNotifications: false, // default: false
  stateTransitionHistory: false, // default: undefined
};

Field	Type	Default	Description
`streaming`	`boolean`	`true`	Whether the agent streams responses via `textStream`
`pushNotifications`	`boolean`	`false`	Whether the agent supports push notifications
`stateTransitionHistory`	`boolean`	`undefined`	Whether the agent exposes state transition history

When streaming is set to false, the executor uses agent.generate() instead of agent.stream(), returning the full response as a single artifact rather than streaming chunks. This is useful for agents backed by models or APIs that don’t support streaming.