The platform

A company, as a system.

Every company here is made of the same few nouns: agents, tasks, runs, decisions, goals and a record. This page is each of them, in the product’s own words and in the states it actually has.

01The workforce

An agent is not a prompt with a name on it.

It has a position, a manager, a voice, standing rules, and an exact set of capabilities: the floor its kind always holds, and whatever you granted on top, one at a time. The org chart is the reporting structure, and it ends at you.

  • Opu cannot be fired

    Your Chief of Staff ships with the company and is always active. It is the top of the chart, and your first point of contact for everything.

  • Two levels, and no departments

    Agents report to Opu, or to an agent who does. Positions are free text: you name the role, the product never forces a template on it.

  • A row, not a redeploy

    Bind a skill, take a tool away, rewrite a rule. The change is live at the agent’s next step. No build, no deploy, no restart.

  • Pause, resume, retire

    Pausing ends every live run and dispatches nothing new, with open work left where it stands. Retiring is the one act that asks you twice.

  • Each agent keeps its own notes

    Short facts an agent chose to remember, listed on its record, each one yours to remove.

  • Talk to any of them

    Opu is where you start. Every other agent is a conversation away, from the button on its own record.

An illustration of an agent’s record: Ivo, in growth, reporting to Opu. It shows the role, voice and rules you set, and a list of tools from the Intercom and Resend connectors, each with a working switch. The reading tools are bound, and so is send_email, which writes and so asks you before each call. create_contact, which also writes, is not bound until you switch it on.

Ivo Active

Growth · reports to Opu

Talk to Ivo

Identity

Edit
Role
Finds the clinics worth talking to, and gets a demo on the calendar.
Voice
Plain and short. Writes like a person who has run a clinic, never like an ad.
Rules
Never emails anyone twice in a week. Never promises a price.
Model
Company standing model
  • search_contactsIntercom
    Read-only
  • get_companyIntercom
    Read-only
  • list_emailsResend
    Read-only
  • create_contactIntercom
    Writes
  • send_emailResend
    Writes

A tool that writes asks you before each run, until you tell its connector otherwise.

Tools

02The work

A task is a contract. Done is a verdict.

A title, a brief, acceptance criteria, what is out of scope, a repository, an owner and a number of attempts. It stands in one of six positions, and a mark says which at a glance. On hold is a flag laid over a position, not a seventh.

acme.opu-inc.com/tasks
TasksLiveTalk to agents
Search tasks or ACM-142Needs you 3Create task

open3

ACM-153

Chase the three unpaid invoices

Sol

Needs you

Reconcile August payouts against Stripe

Capability request: list_payouts

Sol

ACM-152

Draft the churn survey

On hold

unassigned

planned1

ACM-148

Build the inbound lead list for the quarter

Ivo

doing3

ACM-142 /ACM-144

Carry the interval through checkout

Mira

Needs you

Put the annual offer in the upgrade banner

Question: who sees the annual offer?

Mira

Needs you

Write first-touch emails for 14 leads

Approve tool call: send_email

Ivo

verify1

ACM-139

Fix the invoice PDF footer

Mira

done2

ACM-142 /ACM-143

Add the annual price to the pricing table

Mira

ACM-131

Move the docs to the new domain

Mira

cancelled1

ACM-128

Add SMS reminders to the booking flow

Mira

  • open

    Captured, with an owner. No plan yet.

  • planned

    It has a contract: criteria, and a boundary.

  • doing

    A run is executing it right now.

  • verify

    Handed in. A fresh session is judging it.

  • done

    Passed, with evidence. The one state that earns colour.

  • on hold

    A flag, not a position. It waits on you where it stands.

  • cancelled

    It never filled. The record stays.

  • Three runs, three fresh sessions

    Plan writes the contract. Execute does the work in a sandbox of its own. Verify, with no memory of either, tries to prove it failed.

  • Steer it mid-flight

    Interrupt a live run, tell it what to do next, resume it, or end it. A run you stopped waits for you and holds no compute while it does.

  • Three strikes, then you

    An ending that leaves a task no further along costs a strike. At three, the task holds itself and asks you: add attempts, replan, reassign or cancel.

03Goals and standing work

A demo counts when it is on the calendar. Not when an email is sent.

A goal carries a metric with a unit, a target, a direction and a window. A responsibility is the standing work that serves it: one agent, one schedule, one task template.

  • Verified results, never activity

    A metric moves only when a task a responsibility minted passes verification. Emails sent, hours worked and tokens burned move nothing.

  • On a schedule, or on your word

    A responsibility fires on a cron schedule in your company’s time zone, or when you press Run now. Each fire mints a task, under a ceiling you set.

  • A ledger, not a gauge

    Progress is summed from an append-only ledger of outcomes. There is no current value stored anywhere for anyone to edit.

04The record

See what the agent was told. Word for word.

Most tools show you a summary of what the model did. Your company keeps the real thing: the exact text each session ran on, every tool call with its arguments and its answer, and what it cost.

  • The prompt, by segment

    Standing rules, the agent’s brief, the company, the team, the repository. Shown exactly as the model received it, with nothing unattributed hidden.

  • An audit trail by default

    Who asked, on whose word, with what arguments, and what came back. Kept for every call, not for a sample of them.

  • It completes itself

    A run’s ending starts a durable job that finishes its own record. Close the tab, lose your connection, or never watch at all.

  • It stays inside your company

    Prompts and answers are never written to anyone’s telemetry. Your company’s own database holds the record, and its own private store holds whatever a run hands in.

05What it reaches

Discovery is not capability. The binding is.

An agent can see that a tool exists and still be unable to call it. What it can call is decided by one row, the binding, that only you can write.

  • 67 of them

    Tools

    Read, search, edit, run, plan, hand in, judge. Every agent starts with the floor it needs to take part in the whole cycle, and asks for the rest.

  • Installed at runtime

    Skills

    Whole procedures rather than single calls, installed by reference from the public registry. Everything in the package is shown before anything is written.

  • 118 on the roster

    Connectors

    A sign-in on the provider’s own screen, a key, or an installed app. Every operation arrives as its own tool. Reading tools just work, and tools that write ask you first.

  • Any HTTPS MCP server

    Your own servers

    Add a server by its address and its tools join the same list, under the same bindings, with the same rule about writes.

  • 371 models

    Models

    One gateway, every major maker, ready the moment the company exists. Pin a model to any agent, or bring your own key.

  • Console, Slack, WhatsApp

    Channels

    The company comes out of the browser. It admits one person on every channel: you, by the identity that connected it.

Every capability, by name

06What it builds

It doesn’t just use your software. It builds its own.

When the work needs a tool the company lacks, the company builds it: agent-first, with a clean human view over the very same surface. Never one without the other.

  • CRM

    Contacts, companies, pipeline and activity — operated by your agents

  • Tracker

    Tickets, states and owners for whatever the business runs on

  • Support desk

    Customer conversations, triaged and answered

  • Dashboard

    The numbers you actually watch, kept by the agents producing them

  • Held to a standard

    Software built by agents has to be maintained by agents. It is verified like any other task, because nobody is coming to clean up after it.

  • One fabric, not silos

    An app the company builds can read and write another’s data, so the business runs as one operating surface.

07The ground it runs on

Your company isn’t a row in someone else’s database.

It is a deployment of its own, built for you and operated for you. Not a seat in a shared tool, and not a workspace inside a multi-tenant app.

  • Its own everything

    A project, a database, a private file store and an address of its own. Isolation is the architecture here, not a setting.

  • Budgeted before it can spend

    The spend ceiling is set before the deploy that makes spending possible. When it is reached, dispatch pauses and the company tells you.

  • Every session in a sandbox

    A microVM per session, created with no network egress, its push remote disabled. Code runs there, on a branch, and leaves only through a verdict.

  • Releases arrive in rings

    A new release soaks on the platform’s own company first, then a canary, then everyone. An unhealthy company is stepped back and pinned.

The whole isolation story

Name your company. Watch it come into existence.

One name is all it takes. What comes back is a company of your own, at an address of its own, with your Chief of Staff already at their desk.

Create your companyEvery capability, by name