Work / site-agent
An assistant that operates this site in front of you
Ask it something and it flies to the page, points at the section, runs the demo and turns its knobs, then answers with cards built from the site's own content, so it cannot invent a number.
The problem
A site for an agent-systems engineer should show an agent working, not describe one. The obvious version, a chat box that talks about the owner, is the weakest one: it hides the answer behind a text field and it will sooner or later say something that is not true. The site also has to stay a normal site for the visitor who never opens it.
What I built
One stage, several directors. A small client runtime owns the page. It accepts typed actions (navigate, focus, open_project, run_demo, set_param, show_card), validates each one against its schema, animates it and reports the result. It knows nothing about language models. The live model, the recorded tours, the voice mode and an external browser agent all emit the same actions, so they cannot drift apart.
Answers as interface. The agent does not write HTML. It chooses a card from a fixed registry and passes an ID; the card is rendered from the same content files as the pages. A number the visitor sees is a pointer into the content, never generated text.
Two model stacks. Text runs on Claude through one streaming endpoint; the browser runs the tool calls on the page and sends the results back. Voice runs on OpenAI’s Realtime API over WebRTC, speech to speech with function calling, on the same tools and the same prompt. The server mints a short-lived key and never sees the audio.
Guard rails. Per-visitor and daily limits, spend reserved at worst case before each call, a hard daily cap after which the assistant falls back to recorded tours, and a server-signed history so the endpoint cannot be used as a general model. The agent reads only the site’s public content, so it has no private data to leak.
What is not solved
The assistant is as good as the content it reads: a question the site does not answer gets “not on the site” and the contact card, not an answer. Voice depends on a second vendor and costs more per minute than text. The browser-agent path (WebMCP) is a draft standard in an origin trial, so few visitors can use it today.