Agent

An agent is a reusable composition. Harness owns the model-and-tools loop; your application owns the provider client, services, transport, persistence, credentials, and deployment.

import { Agent, tool } from "@nylorun/harness";
import { z } from "zod";

const echo = tool({
  name: "echo",
  description: "Return the supplied text.",
  parameters: z.object({ text: z.string() }),
  execute: async ({ text }) => ({ kind: "completed", output: text }),
});

const agent = Agent({
  id: "support",
  name: "Support",
  instructions: "Be concise and use tools when they help.",
})
  .use({ id: "support-tools", tools: [echo] })
  .with(adapter)
  .build();

id is a stable programmatic identifier and name is a display name. instructions may be one string or an ordered list of strings.

Compose with capabilities

Pass a CapabilityDeclaration to .use() to attach model-visible instructions, tools, a model directive, optional per-session state, and optional middleware under one stable id. The declaration keeps each concern attributable in the resulting model configuration and observations.

Use separate capabilities for independent application services—for example, a search client, a customer record store, or approval policy. Construct those services in your host application and close over them in the capability; Harness does not construct or configure them for you.

Bind once, then build

.with(adapter) binds exactly one ModelAdapter and returns a bound builder. A bound builder only exposes .build(), which prevents later mutation of the composition. build() is idempotent: a successful builder returns the same BuiltAgent instance on subsequent calls.

Next

Choose and connect a provider in Model, then add policy and session behavior with Middleware and Session.