With this, your mind can give its members a home for their agents: a first-message box, recent work, and a chat that stays with each agent’s durable log. This is a privileged workbench, not a chatbot exposed to visitors. Starting an ordinary agent gives it the mind’s reach; moving the chat into a pretty page does not make that safe for strangers.

A small host, a reusable UI

apps/agents/app.ts installs the Agents app through installAgents from iterate/agents/install, pointing it at apps/agents/backend.ts. The backend re-exports the platform’s agent classes and chooses existing model funding and standing instruction files. index.html pins browser package versions; public/app.js wires the catalog, chat and agent kinds; public/home.js handles the first message and recent rows. Navigation and row helpers are separate local files rather than duplicated behavior in each view. Home cards and sidebar rows also show a working or waiting agent’s current activity below its title. On narrow screens those summaries can wrap rather than forcing a wider page; pins remain a per-browser navigation choice, not a durable authorization setting.

import { agentsApp } from "./apps/agents/app.ts";
const agents = agentsApp();
// Run agents.install(itx) in the idempotent publish hook.
case "agents":
  return agents.open?.(request) ||
    this.auth.require(request) || agents.fetch(request, this);

The open exception serves a narrowly allowlisted icon or app manifest, not chats or the member socket. The guarded /itx route calls host.auth.itx(request). The browser opens a WebSocket on that same host and wraps it with the UI kit’s capnweb session helper. The platform decides the principal and ends the socket when the sign-in ends; browser-supplied identity is not trusted. Three compatibility endpoints describe the one mind, the session issuer and sign-out for the UI shell.

Agents are contexts with a kind

The catalog and chat are backed by agent contexts under /agents, not a second database inside the page. The first message creates an agent and sends a message through its facet. The ordinary iterate kind uses the platform runtime; other kind modules offer coding runners or different views. agent-kinds/ files name their kind, optional start choices, event processing and view. worker.ts registers the kinds alongside integrations. Private model and tool events remain within the member boundary; this introduction contains no conversation examples from real work.

Give the catalog names, not old tool output

Each runner's kind module can fold its events into agent/listing-updated facts on the /agents catalog. Current Claude, Codex, ACP and Herdr modules also copy agent/summary-updated: an agent's own nonempty title, description and activity can name its enduring role separately from its current work. Named summaries take precedence over fallback titles derived from the first words of a request. Fallbacks strip common relay or brief prefixes and keep a short phrase; they are labels, not instructions or evidence of identity.

Historical summary rows can rebuild metadata from the durable log. Claude and ACP tool-activity rows instead start from now, so replaying setup does not present old tool descriptions as live work. A person's new request clears prior activity where the runner supplies that transition. Keep status, role and current activity as separate fields; test a named agent, an unnamed session, summary replay and a new request after an idle turn. Rebuilding a catalog should not restart the underlying work.

Copy the boundary as well as the files

A similar app needs the iterate agent runtime, the UI kit at a tested pinned version, and existing model funding: a suitable connected plan, API key or gateway credit. A different runner can need its own sandbox or computer. None of this requires transferring our subscriptions or secrets. Preserve the guarded socket and route every chat action as the signed-in member. A public sales or install conversation needs a separate restricted backend, not this ordinary workbench. Coding agents describes the runner side; onboarding shows a narrower member page using one fixed agent; the blog records changes as recipes.

Verify the workbench

Test an empty catalog, creating an agent from a first message, returning to its durable chat, switching kinds and reconnecting after a socket closes. Check that signed-out visitors and signed-in nonmembers cannot reach the socket directly. Exercise long titles and working summaries on narrow and desktop layouts. Catalog status tells a member what a runner last reported; it is not proof that a remote process is still connected. Keep errors visible instead of silently sending a message to a different runner.