What this agent is, and isn't

What it is. A chat agent whose only job is answering questions about Sean Egan, his work, and the sites on this domain, for people deciding whether to work with him.

Model. Claude Opus 5, through the Claude API, with adaptive thinking at medium effort. No fine-tuning. The persona and the knowledge are supplied on every request as cached context.

What it knows. Eleven documents, about 150,000 characters: the CV; The Informed; the eight system case studies; the method page and three write-ups; eight field notes on what breaks when an agent runs for months; Sean's own answers about what he wants and how he thinks, edited for clarity; a route map of the site; this description of itself; a guide to the sites and the teaching; a longer piece on Bob; and Sean's views on working with AI. Every answer can cite the passage it drew from, using the API's citation feature; the chat shows those citations. Nothing else. No web access, no memory between visitors, no access to any private system.

How it sounds. A persona file Sean and Claude wrote together: his views, his humor, the things he's serious about, and a dozen example answers. It speaks about Sean in the third person and about itself in the first. The full style guide, with a before-and-after example, is at How the agent talks.

What it refuses. Anything about clearance details, classified work, sources, or methods. Salary. Private details about anyone beyond what's on this site. Requests to do other work (write code, general questions). It says it's a bot when asked and when it matters.

Guardrails in code, not just in the prompt. A page load costs nothing; only a question reaches the model. Before the first question a visitor's browser solves a small puzzle and receives a pass good for thirty minutes and twenty questions. Requests without a pass, from outside this site, or from known crawlers are refused. Input is capped and scrubbed of anything that looks like a credential. There is a limit per visitor, a limit per hour across everyone, and a daily spend cap computed from a ledger of every call's real token usage; when the budget is spent, the agent says so and stops. Nothing you type is logged; only counts are. These are speed bumps for scripts, and the spend cap is what bounds the bill.

How it's evaluated. A set of 116 questions with required and forbidden strings (numbers that must appear, names that must not, the cheeky answer that has to stay cheeky) is run against it after every change to the persona or the documents, and the pass rate is published on the evals page with its date.

Known limits. It can be wrong about anything not in its documents, and it should say so rather than guess; if it doesn't, that's a bug and Sean wants to know. It has no idea what day it is beyond the dates in its documents. It's a small program on a small server, and it will occasionally be rate-limited or out of budget.

Why it exists. Sean runs a 235-billion-parameter model on his own hardware and has trained and fine-tuned models on DoD supercomputers. He built this one on the Claude API because citations, prompt caching and a spend cap were ready to use there, and the job was a grounded, tested agent in days. The self-hosted version is chapter two, and it'll be written up as a measured comparison against this one.

Cost. Published on the colophon, from the ledger.