Runner
Sign in

What an agent is, and where every answer comes from

An agent is a face, a job, and a briefing assembled from records you already hold. This page is the part the landing page does not have room for: where the briefing comes from, what happens when an agent has not been told something, and where a visitor actually meets one.

How one is made

  1. 01Pick a characterEvery character Runner ships is rendered for this and owned outright. None is a photograph, none is a likeness of anyone, and none has a name — a role only. Whichever you pick carries the AI mark wherever it appears — nothing in the product removes it, so a visitor is never in doubt.
  2. 02Give it a jobThe role decides what the agent is briefed on and which records it may read. Six roles arrive seeded, and an organization can define its own. The role is what scopes the briefing, so it is the choice that matters most.
  3. 03It briefs itselfThe briefing is assembled from your workspace's own records, scoped to your organization and nobody else's. You do not write it, and you do not maintain it: when the underlying records change, the next briefing is built from what is there then.

What Runner keeps in house

Runner owns the character, the role system, the briefing assembly, the evidence check, the AI label and the handoff to a person. Those parts stay in Runner, beside the records they govern.

Connect the model account your team already uses to generate the conversation. That account can improve the answer; it does not widen the records the agent may read or remove the citation rule.

It cites, or it says it does not know

Every claim you produce MUST cite exactly one supplied evidence record by its id, with a verbatim excerpt from that record. A statement you cannot cite is a statement you must not make.

The evidence rule carried by every agent briefing

Each claim an agent makes is tied to one record it can point at, with the words it took from it. That is the whole mechanism, and it is the same one for every agent in every role.

A briefing is checked before it is stored. If any claim in it points at a record the agent was not given, the briefing is refused — not trimmed, not saved with that claim removed. Nothing is stored at all.

The other half is what happens when there is no such record. The agent does not soften the claim or reach for something adjacent — it reports that it has not been told, and a person picks it up. An agent that has run out of evidence is supposed to be visibly out of evidence.

The evidence an agent may read is scoped to your organization. There is no shared pool between workspaces, and nothing an agent learns in yours is readable from anyone else's.

Every answer an agent gives is built the same way, in the same order — and the last step is the one worth knowing about.

  1. It starts with your business

    Only what your team has already written down, and only the current version.

  2. It forgets everything else

    Whatever it picked up elsewhere does not count here.

  3. It shows its working

    Every line points back to where it came from, in your own words.

  4. Then we check it

    Nothing is taken on the agent's word, including by us.

  5. Or it tells you it does not know

    Nothing to point at, nothing said. It would rather stop than guess.

This is not a setting somebody remembered to switch on. It is how every agent answers, every time.

CADENCE — what one sounds like

AI-generated face

Technology director

CADENCE gives each character its own voice — Runner's catalogue rather than a vendor's, some of them blended from two packs, each paced to how that person would actually speak. Press to hear one. It plays when you ask it to and not before.

Plays once, with sound. Nothing starts on its own.

Runner TRACK — how one runs

AI-generated face

Runner TRACK
STRIDE
0000 / 000
0.00s
SPLIT
▯▯▯▯▯
SIGHTLINE
—
FORM
technology-director

Live frame readout

Runner TRACK 1.3

Every agent on this wall runs on Runner TRACK, our own live-agent stack. Not a stock avatar anywhere — each is a synthetic person we built, rendered and voiced on our own infrastructure.

  • STRIDEMoves the face — a whole, seamless loop from one still.
  • CADENCEGives each one a voice, paced the way that person would speak.
  • SIGHTLINEReads the eyes off the frames, so the overlay reports them.
  • SPLITReads the mouth the same way, frame by frame.
  • FORMHolds the identity each agent keeps wherever it appears.

Version 1.3, live on this wall right now. The next one is already being built.

The cast Runner ships

Every one is synthetic and owned outright — not a photograph, not a likeness of anyone, and never given a name. The role is the caption, and the mark is on every face. Your team can add characters of its own alongside them.

  • AI-generated face

    Technology director

  • AI-generated face

    Product manager

  • AI-generated face

    Sales AE

  • AI-generated face

    SDR

  • AI-generated face

    Customer success

  • AI-generated face

    Recruiter

  • AI-generated face

    Onboarding guide

  • AI-generated face

    Research analyst

Where a visitor meets one

On your site, an agent appears as a portrait beside a text conversation. It says it is an AI, and one tap asks for a person: a teammate is notified and can take the conversation over from the same record.

A two-way video call with a visitor is not offered. The talking takes on this page are agents an operator has built, running on the operator's own calls — not a stream to a stranger on your website.