Agentcy Core brings client context, named agents, and accountable follow-through into the agency’s working day.

The conversational surface where sessions happen. Slack is an outbound destination for supported messages and configured nudges, not the place the work is done.

Reviewed account knowledge. Alongside it, the Review Queue holds facts staged for a human decision, and the Audit View holds the record of what was decided and produced.

HTML work served and shared in the browser. A page a seat composes renders in the Account’s brand profile — palette, fonts, logo, tagline — or in the house brand an Admin set, if the Account has none. An Account’s own fields win. A recapped call can be made into a tearsheet from the Desk by an Editor. Its provenance — the facts it drew on and the Gate’s verdicts — stays on its record in the workspace; a shared link shows the page, not the receipts. Link creation is gated; links expire, can be revoked, and can carry a password. Revocation stops future access — it cannot erase a copy someone already made. From a tearsheet, Send on the inspect card briefs the Art Director on that Account with the element you captured and the instruction you typed. It does not brief the Supervisor. A designed page can be downloaded as a file. The platform does not host it, deploy it, or publish a URL for it.

The calls the meeting bot sat in for you: the title, when, which Account it sits on, and where the transcript went — into that Account’s review queue as facts, and to the seat that sent the bot as a recap. A call that matched no Account is never guessed at by a model. Filing one by hand takes an Admin, for calls they sent, onto an Account or onto Agency. Agency is the agency book, for partners and vendors. It is not a client, and it is not a guess.

Recurring work an Admin approves once, acting for a named person: a check that runs on a schedule and only wakes a seat when there is something to see, or a turn that runs on its own. The schedule is a small enum read back in plain English, never a cron line.

Each morning at nine, the agency’s own time, the Chief of Staff writes the day’s brief from the agency’s live record and proposes it like any other draft: gated, then decided with Approve or Reject on the front page. The format does not accept an uncited claim, and a morning where a source refuses still produces a brief that names the gap rather than staying silent.

News about one Account: the Account itself, up to three competitors, up to three platforms, and up to two topics. Opening the page reads cards already saved. It does not search, and a card is not a fact and does not enter the Book. An Editor names the list and runs the search. A daily Standing Order, proposed once by the Chief of Staff, searches at 8:00 Pacific for each Account that has a list. An empty list is not searched. Pausing the order is the off switch.

Seats propose, people ratify. What a failed connected-app call taught is recorded with the receipt it was learned from, and what a seat learns about doing its role well can be proposed as craft shared with every seat holding that role — neither changes behaviour until a person adopts it in Admin. An adopted lesson is advice, never a wall: what is actually connected is re-read live on every call, and a person asking again outranks a remembered refusal.

Cards the seats raise: one observation, why it matters now, and a suggested action, with the evidence attached. An Editor applies or dismisses a card, or hands it back to the seat, which does the work and writes the outcome onto the card.

Playbooks are procedures a person authors and versions in Admin; a seat runs the version it was handed and says so. Tickets are what a seat files when a capability is missing or a service refused — deduplicated, kept on your own instance rather than sent anywhere, and answered by a person in Admin.

A question in the top bar, answered from the written pages, or it says the guide does not cover that. It is not a seat, and it is not the Desk.

The workspace ledger: what is waiting on you, what was filed, which facts have gone stale, which sessions are open. Accounts are client records inside your agency’s tenant. They are not client logins, and there is no shipped client portal.

PDF, DOCX, PPTX, XLSX, Markdown, plain text, VTT transcripts, images, files pulled from Google Drive or Box, attachments from the person’s Gmail — which a Supervisor can file into the Book for review — and the transcript of a recorded meeting, read with source locators, so a passage can be traced to a page, a slide, a section, or the stretch of a transcript where the speaker is named. Superhuman docs, Canva designs, and Notion pages the person has shared with the integration can be read in a conversation as working material; they are not filed into the Book.

A vision path transcribes image-only PDFs that are uploaded, attached on the Desk, or filed into the Book. It runs as a fallback, only when a file yields no extractable text; a seat’s first look at a scanned file does not use it. A loose image reads as a one-page scan, and a diagram with no words on it — a flow sketch, a floor plan — can have its described interpretation staged for review, marked as an interpretation rather than a transcription.

25 MB per file. A document over 300 pages has to be split before it can be filed into the Book; a seat can still read a longer document in the conversation up to a bounded length, and the reading says where it stops.

Extracted Account facts enter staging, then a human review decision, and only then live retrieval. Rejection is a normal outcome and is recorded.

A critic scores every staged fact and names the keeper in a cluster of near-duplicates; the rest retire with it on one click. A live fact leaves the Book only by a recorded retraction or by naming the fact that replaces it.

Named execute-class operations a person asks for in their own conversation are gated and recorded, and run without a second approval card. You already asked.

Named execute-class work raised by a fired timer waits as a proposal until someone approves it on the Desk, unless the person replies in that conversation, which makes it theirs. It surfaces under “Needs you,” with a Slack nudge where that is configured. Changes to a Box file and code runs are refused instead, and a draft email, a task or an invitation needs that person’s yes in Admin → Delivery. Work the Chief of Staff hands to a Supervisor from that conversation does not wait.

A Standing Order acts for one named person: the person whose conversation it was proposed in or, if there was none, the Admin who approved it. It is approved once, not on every run: named execute-class work in its runs does not wait for a card, while draft emails, tasks and invitations go only where that person has said yes in Admin → Delivery. A check-kind order looks and stays quiet when there is nothing to see — one ledger row, no session, no tokens. Nothing runs as “the agency.” If its seat is retired, the next run fails and the order pauses; when a person leaves, an Admin pauses, reassigns or cancels their orders — that step is not yet automatic.

Among the named operations, reading and drafting are not writing. An email draft stays unsent in the drafts folder. A document comment is a real write, made under the identity of the person whose connection was used. The same is not true of tools a seat finds by looking through a connected app.

Deterministic rules run against configured policy and return a versioned verdict. Rules that need a model are not evaluated yet; the Gate lists each one as skipped rather than passing it quietly.

Artifact records, retrieval snapshots, review decisions, policy versions, and action receipts are inspectable. That is an operational record. It is not regulatory compliance, insurance eligibility, or an audit attestation.

One instance and one database per agency. Inside it, each Account seat — the Supervisor and the Art Director — is scoped to its Account; people in the agency are not — anyone with a role can see every Account. Seat proofs bind service access to stored identity and account assignment rather than to a scope the model asserts about itself.

Account boundaries govern the account record. They do not partition a person’s own connected Drive or calendar, which the person already has access to.

A hosted orchestration runtime with durable session and turn storage, tool dispatch, cancellation, budgets, and checkpoint resumption. Adapters exist for Anthropic, OpenAI, Google and xAI; the hosted process runs one selected backend rather than switching models per turn.

The conversation model is the agency’s own. Connect a Grok (SuperGrok) subscription from Admin — a short code, approved at xAI on the account that holds the plan, picked up by the running instance within a minute — or store an xAI console key or a Z.ai GLM Coding Plan key, verified against the provider before it is stored and held encrypted under a key that never enters the database. Your plan or your meter pays for your seats’ thinking; platform work keeps its own keys, below.

Conversation turns go to the one backend the instance runs. Document intake, scanned pages and meeting recaps use a platform-funded Anthropic lane; the embedding path uses OpenAI; calls are recorded and transcribed by a meeting-recording service; image generation for the Art Director uses a platform key; and most connected-app calls pass through a connector service that holds each person’s app authorizations. “Only your own model provider sees your data” would therefore be untrue, so it is not claimed.