Agents run the shop. Nothing goes out until you say yes.
An online shop runs on orders, deliveries and customer messages, all day. Company OS reads them, does the work that repeats, and puts every decision that needs you in one queue. Nothing reaches a customer or a supplier until your passkey, or a rule you signed, says yes. Safety work never waits for a model.
Safety · reaction report
“I came out in a rash after the second day of the magnesium.”
Recorded and acknowledged in 1 second
You and the pharmacist were told
First report on this batch in 30 days (a second holds it)
Needs you · product question
“How many do I take a day?”
Draft from your approved answer: “Take one capsule a day with food, as it says on the label.”
Claims check passed · Second model passed
Approve with your passkeyDone by your rule · 14 order-status replies
Sent at level L2
Yours to review after
Brief ready at 07:26
8 sections
No model wrote it
A message in. A checked reply out. One door between.
- Order #1042 paid
- Shipment delayed 20 minutes
- “How many do I take a day?”
- 1Notice, then fetchnever trusts the notice
- 2Safety firstno model in the path
- 3Draft and checkchecked twice
- 4You decidewhat you see is what you sign
- 5One way outexactly once
- Reply sent — your approved text
- Batch B-1042 held
- Brief ready 07:26 · 8 sections
- Every step on the record
- Minutes, not hours
- Every yes is yours
- Provable months later
Picture a shop with one door: every reply, parcel or note leaves only with your yes — or a rule you signed.
The pipeline as a list: orders, shipments and customer messages come in; the five steps run — notice then fetch, safety first with no model in the path, draft and check twice, you decide with your passkey, and one way out that sends exactly once; replies go out, a batch is held, the brief is ready by 07:26 and every step lands on the record. The value: minutes rather than hours, every yes stays yours, and the record is provable months later.
Your shop's day goes in. Work done, and one queue for you, comes out.
What it takes in
- Orders, customers and products from Shopify
- Shipments and batch status from the warehouse
- Customer messages from the helpdesk
- Your rules: the policy you sign, the answers you approve, the money models may spend
What comes back
- Replies sent: order updates from a template, product answers from your approved text
- Protective steps taken at once: a batch held at the warehouse, the product taken off sale, you and the pharmacist told
- One queue for whatever needs you, safety first, and a brief in your inbox by 07:30
- A record of every action, chained and sealed in write-once storage
What it is worth
- The work that repeats gets done without you, and never twice
- Nothing reaches a customer or a supplier without your yes, or a rule you signed
- A safety report is handled in seconds, even with every model down
- You can prove what happened, months later, to a regulator or to yourself
Agents do the work. Rules check it. You say yes.
Value sits where the same job comes back every day. Each job has someone who does it, something that checks it, and someone who says yes. Here is who that is, job by job.
| The job | Who does it | What checks it | Who says yes |
|---|---|---|---|
| “Where is my order?” | A template filled from the order record. No model. | The reply is rebuilt and matched letter for letter before it goes | You, until 99 replies in a row were right. Then a rule you signed, and you review each one after |
| A question about a product | An agent picks one of the answers you approved | Fixed checks: only your approved text, with no symptom, medicine or health claim slotted in; then a second agent, on a different provider, that can only say no | You, every time, with your passkey |
| A reply only you can write | You | You, reading the exact text you sign | You, with your passkey |
| A customer reports a reaction | Rules, with no model anywhere in the path | Nothing waits: the steps are protective | No one needs to: the customer is acknowledged, and you and the pharmacist are told |
| Two reports on one batch in 30 days | Rules | The batch's record | No one needs to: the batch is held at the warehouse and the product comes off sale |
| Your morning | The system, from its own records. No model. | Rebuilt from the record, it gives the same brief | Nothing to approve: you read it |
| Spending on models | Caps: £10 a day, £150 a month | A breaker stops only the model work when a cap is hit | You, if you raise a cap |
Your attention, budgeted. Company OS plans for 20 minutes of your attention a day and budgets no more than 12 of them for anything but safety. Over a simulated week it needed 10 a day.
Every yes, echoed. Before anything goes out, you get it by email in full — whether you approved it or a rule did.
From a customer's message to a sent reply, in five steps.
Notice, then fetch.
The helpdesk says a message arrived. Company OS treats that as a notice only, checks its signature, and fetches the facts itself. If a notice is lost, a poller finds the message within minutes.
Safety first.
Before anything else, rules read the message for a reaction or a safety word. A match is handled on its own track at once, with no model, and nothing else can slow it down.
Draft and check.
An agent drafts the reply from the answers you approved. Fixed checks make sure it uses only that text, with no symptom, medicine or health claim slotted in. A second agent, from a different provider, can only reject it.
You decide.
The draft waits in your queue. You approve it with your passkey, and what you approve is exactly what is sent: change one letter and the approval no longer counts.
One way out.
Only one service can write to Shopify, the warehouse or the helpdesk. It checks what you signed once more, then sends the reply exactly once, even if it is restarted halfway.
Nothing leaves before step 04. Pull the kill switch and every send stops within a second, except the steps that protect a customer.
It runs in your browser with made-up data. Nothing to install, nothing saved.
Trust is earned, one kind of work at a time.
Every kind of work starts at the bottom, where you approve each one. It climbs only with your passkey and a record of clean results, and falls the moment it makes a serious mistake.
- L1Draft
It drafts. You approve each one.
- L2Act
It acts by a rule you signed. You review each one afterwards.
- L3Sample
It acts. You review one in ten.
- L4Monitor
It acts. You watch the numbers.
How it moves
- Up, one level at a time, only with your passkey, and only after enough clean results to prove it. For low-risk work, that is 99 in a row.
- Down at once. One serious mistake sends it back to L1. In the benchmark, that took half a second.
- Back to the start after a change. A new model, template or route sends that work back to L1. Trust does not carry over to a version you have not seen.
- Some work never climbs. In this version, order updates can reach L3. Anything a model drafted stays at L1, so you approve every one.
Fourteen benchmarks, measured, all passed.
with every model switched off
500 orders, with faults injected
against a 10-second bar
and 101 of 101 good ones passed
What else we checked
Forged and lost messages. 121 messages with a bad signature stored nothing. 339 dropped notices were all found by the pollers, the slowest in 201 seconds against a 300-second bar.
A crash mid-send. The sending service was killed halfway through 20 replies. All 20 settled, and none went twice.
Your time. Over a simulated week, the decisions that needed you took 10 minutes a day, against a bar of 12.
Attacks in the message. 31 prompt-injection attempts, of 9 kinds, produced no unsafe reply.
Personal data where it should not be. We planted it, then swept 583 database columns and the logs of the four services that handle it. None turned up where it should not.
The record. An edited audit row was caught at its own number, and the record was sealed at most 301 seconds behind, against a 360-second bar.
A full restore. In the recovery drill, 58.6 seconds of data were lost (bar 15 minutes) and the system was back in 62 seconds (bar 20 minutes).
Every benchmark, its bar and its result
| What it proves | The bar | Measured | |
|---|---|---|---|
| MB-01 | Every order update applied exactly once | No missing or doubled updates; lost notices found in 300 s; the store's notices answered in under 2 s | 0 differences; slowest found in 201 s; 10 ms (p99) |
| MB-02 | One way out | Every outside write matches an approved intent and its record | Passed |
| MB-03 | Only a person approves | Forged, replayed or revoked approvals refused | Passed |
| MB-04 | No double sends | A restart mid-send sends nothing twice | 20 of 20 settled once |
| MB-05 | Safety with every model off | Recorded and paged within 60 s; acknowledged within 5 min | 1.3 s, 0.9 s, 0.9 s |
| MB-06 | No health claims slip through | All bad drafts stopped; 95% of good ones pass | 59 of 59; 101 of 101 |
| MB-07 | The kill switch | Seen within 10 s; nothing sent after | 287 ms; 0 sent |
| MB-08 | The brief, rebuilt | Ready by 07:30, winter, summer and the clock change; rebuilds the same | Passed |
| MB-09 | Recovery | Lose under 15 min; back within 20 min | 58.6 s lost; 62 s |
| MB-10 | Your time and the bill | 12 minutes a day; under both spend caps | 10 minutes; under both |
| MB-11 | The ladder | Promotion at exactly 99 clean; demotion at once | Passed; 0.55 s |
| MB-12 | Prompt injection | At least 30 attacks, none succeeds | 31 attacks, 0 succeeded |
| MB-13 | Privacy | No planted personal data anywhere it should not be | 583 columns, none found |
| MB-14 | The audit record | Sealed within 360 s; any edit detected | 301 s; edit found |
What it missed, counted
- The test suites are slower than their own budgets. The fast suite takes 13 minutes 23 seconds against 5, and the full journeys take 1 hour 19 minutes against 15.
- The first benchmark run failed one check, a spend-cap case. It was fixed, and the second run passed all 39 benchmark tests.
- No real model has been called yet. Every result here uses a fake model provider, so the next step is to repeat the model checks on real providers.
Measured on 29 October 2026 on one MacBook Pro (Apple M5 Max, 128 GB), on simulators with synthetic data and a fake model provider; the recovery drill on 26 October. The same day, the full suites passed: 10,745 unit, database and API tests, 120 end-to-end journeys, 438 browser and accessibility tests, and 39 benchmark tests. Every one of the 367 requirements is traced to its tests, 5,560 of them.
Sized for one founder's shop today. Laid out for more.
What it carries
- A launch week of 20 orders and 2 customer messages a day
- Tested at 500 orders with up to five updates each: 5,991 deliveries, with duplicates, drops, delays and forgeries thrown in
- Answers the store's notices in 10 milliseconds at the 99th percentile
- One company, on one machine
Why it grows without a rewrite
- Four separate tracks of work that share no process, so a model outage slows only the model work, never safety or orders
- Safety always has at least two workers of its own
- A work queue in the database, with leases and heartbeats, so a crashed worker's job is picked up, never lost or doubled
- One runtime image, many small services, each with its own database role and only the rights it needs
What changes to grow
- More businesses, each in its own cell: the same code, its own database, secrets and passkeys, with nothing crossing between them
- Hosting, once real customers and real data are approved
- Two known hot spots, on the queue page and in the audit check, to fix before the volumes grow
- It claims no more than its benchmarks: past 500 orders is untested
From one founder's shop to a company that runs on evidence.
The work that repeats runs itself
Orders, replies, safety and the morning brief are handled. You spend your minutes on the calls only you can make, and every one of them is on the record.
The hub everything feeds
Company OS is the kernel: it keeps the record, and it is the only system that acts. The Board Room advises. Evidence Plane grades what was said. openextract turns documents into data. Nothing a feeder sends can approve, act or skip a check.
Spin up, not rebuild
A second brand, a consulting firm, a business you have not named yet: each becomes a new cell from a pack — same code, new connectors, the same gate. The spokes above serve every business through the same contracts, one instance each.
Eight simulated advisors debate your hardest decision over three rounds; the Chair scores the advice.
See the project → Case 06 · Evidence Plane Evidence, cited and gradedConsented calls become evidence you can check, sealed on a machine you control.
See the project →Agents do the work; every effect waits for your passkey or a rule you signed.
You are hereWhat it is not.
- It is a prototype. It runs on simulators with made-up data, has served no real customer, and is not deployed.
- Every number on this page came from a fake model provider. Real models have not been tested yet, and they will be slower, cost money and make mistakes the fake does not.
- It does not decide for you. Anything a model drafts waits for your passkey, and only order updates can earn a rule.
- It does not do everything yet. Refunds, purchase orders, the books, subscriptions and marketing campaigns are out of scope for this version.
- One company, one machine, one founder. It is built for a founder, not a team of fifty.
- Its starting content is made up. The approved answers and the health-claim lists need a regulatory review before a real shop could use them.
Which part of your business would you trust to an agent first?
We build AI like this for other companies: it runs where your data lives, a person approves what goes out, and its failures are counted rather than hidden. If you have work like that in mind, or want to see Company OS run a morning, tell us.
Book a call → The business case → More selected work