The founder
I built this because I watched people get let down by confident answers.
Forge Agents is one person in Minnesota. I write the code, I answer the email, and I decide what this software will and will not do. This page is the last one, because I would rather you judged the product first.
Where it starts
In the Army, a guess with confidence behind it got people hurt.
That is the first place I learned the difference between an answer and an assurance. You are taught to say "I don't know, I'll find out" — out loud, in front of people, without flinching — because the alternative costs more than your pride does. It is an uncomfortable sentence the first hundred times. Then it becomes the most useful one you own.
I carried that into network administration, where the job is mostly the same lesson wearing different clothes. Every serious outage I have worked started with somebody who was sure. Not malicious. Not lazy. Sure. The systems that survive are the ones that make it easy to say "that isn't documented" and hard to invent something that sounds right.
What I kept seeing
The person at the front desk is holding the whole building up.
Walk into any small practice and watch the front desk for an hour. Someone is on the phone, someone else is standing at the counter, and a third person is asking a question that has a documented answer sitting in a binder eleven feet away. So the answer comes from memory. Usually memory is right. When it is wrong, it is wrong quietly — and the person who gets blamed is the one who was trying hardest to help.
Nobody designs that failure. It is what happens when the knowledge lives in a binder, a shared drive, and one person's head, and the phone does not stop ringing.
The conviction
An assistant that guesses is worse than no assistant at all.
Not equal to. Worse. A person who has no answer goes and finds one. A person handed a confident wrong answer acts on it — and every safeguard downstream is now working against a thing that looks like it came from the documentation. That is the whole reason this product refuses to be clever. It quotes your procedures or it tells you it cannot.
How that shows up
Every choice here is downstream of one refusal.
The software checks its own citations against your documents before you ever see the answer, and drops any it cannot prove. It tells you when it has nothing rather than filling the silence. It will not give clinical advice at any volume of pressure. The price is printed on the website because making you sit through a call to learn a number is a small dishonesty, and small dishonesties are the whole thing I am trying to build away from.
None of that is a feature list. It is the same decision, made over and over, in the places where a shortcut would have been easier.
Why it stays small
I would rather be trusted by a few offices than known by many.
Forge is deliberately narrow. One product that does one job for one kind of office, proven before it is sold anywhere else. If we ever get big, I want it to be because the thing worked, not because the pitch did.
If you run a practice and you want to see whether this is real, ask it something your documents do not cover. Watch what it does with the silence. That is the part I am proud of.
For my fellow builders who care about what's inside
The refusal you just read about is not a system prompt asking nicely. Every citation is machine-verified verbatim against the source corpus after generation; an answer that cannot survive that check doesn't ship, and a claim the documents don't contain returns as a first-class refusal with a route to a human. The multi-part rule is enforced the same way — answer half a question, flag the other half as undocumented, never let a partial read as complete.
We build from the literature, not vibes. Recent agent-systems research is implemented here, not summarized: execution-validated task synthesis generates each customer's acceptance suite from their own corpus and gates it on execution — grounding, vacuity, adversarial wrong-answer checks — before anything goes live. Memory-trap batteries pressure-test refusal integrity across multi-turn emotional, false-authority, and cross-language attacks; that work took our trap suite from mostly-passing to clean. Prompt evolution runs as variation operators under a fixed test oracle — the tests never bend to the model.
The evals get evals: our test suites are themselves linted for false-fail constructions before any model run. Corpus changes are insertion-only with byte-exact provenance gates, so a customer's documents can be deepened but never silently rewritten. The connector is a stateless per-request server behind OAuth 2.1 — dynamic registration, PKCE, rotating refresh tokens with family revocation on replay — and its auth layer mirrors the identity store read-only: it cannot write, by construction, not by policy. Deploys are health-gated with automatic rollback.
The honest version: the parts that make it refuse are tested harder than the parts that make it answer. That is the whole company.