One window cleaner, three hundred customers: a solo round's back office.

A one-person window round runs on a weekly plan that the weather rewrites. In the illustration below, a routine prepares the week before the cleaner sets off, a Bot proposes a new order of streets when rain stops work, and the cleaner approves or rejects it from a phone between houses. Another Bot reconciles cleans against payments and drafts the chasers. None of them decides whether to work in the rain, and none of them drops a customer who will not pay: that decision is the cleaner's, every time.

This round is made up to show the design: no real firm stands behind it, nothing here is a result, and the Bots in it are ones the cleaner creates under Aevrion's pattern, not assistants that Grok Bot supplies ready-made.

Based on the Grok Bot documentation at docs.x.ai as checked on 15 October 2026. Its statements on Bot rules, Bot-to-Bot messaging, routine triggers, connected apps, approval expiry and payments were rechecked against docs.x.ai and Cursor's help pages on 26 October 2026. The trade examples are Aevrion's illustrations, not documented Grok Bot use cases. Aevrion Ops is an independent consultancy. It is not affiliated with, endorsed by or a partner of xAI, SpaceXSI or Cursor. Grok Bot is named here only to describe it.

The cleaner in this example works alone from a van carrying purified water and a water-fed pole, which lets nearly every window on the round be cleaned from the ground. That matches the first step on HSE's window cleaning page (read on 16 October 2026): avoid work at height where it is reasonably practicable, with water-fed poles given as an example. A few addresses need something else, and the choice of method there is the cleaner's alone. HSE also notes that health and safety law applies to a self-employed window cleaner whose work exposes other people to risk, which on a pavement full of hose is not a remote idea.

About two hundred customers are on a four-weekly cycle and the rest on eight-weekly, kept in the round-management software the cleaner already pays for. Customers pay in different ways: some by card on file, some by bank transfer after a note through the door, a few in cash on the next visit. Quoting, the round order, reminders and replies all happen in the evening, after the van is unloaded.

The One-Man Army model, set out at length in its pillar article, keeps every decision with the person running the business and gives each repeating piece of paperwork to its own named assistant. Here that means three Bots, a small group for them to work in, and a cleaner who stays the only person who approves anything.

The documentation suggests a new Bot only when the work brings its own aim, tools, sign-off point or timetable, and a roster kept as small as it can be. On this round that gives three, and the cleaner added them one at a time.

• ROUND PLANNER

  Builds the week's order of streets from the next-due list, re-proposes it when work stops, and coordinates the group. It never decides that a day is lost or applies a plan the cleaner has not approved.

• COLLECTIONS

  Matches the cleans marked done against the payments in the cleaner's bank export and drafts the chasers. It never takes a payment and never removes anyone from the round.

• CUSTOMER NOTES

  Drafts the emails that go with a moved clean, a skipped house or a new customer's first visit, and sorts the replies the cleaner passes to it. It never promises a date the Round Planner has not proposed and the cleaner has not accepted.

The week below is one illustrative week on this round. Each day names the decision the cleaner makes and what exists once it is made. The Bots read the round software through a custom integration, built and tested against what the vendor's API or terms allow, or by supervised use of the vendor's website where its terms allow; if neither is possible, the cleaner exports the due list and drops it in the shared workspace.

1. MON

   The prep routine

   Decides

   Before setting off, the cleaner reads the week's plan and decides whether to work it as written or move a street himself.

   Output

   A plan file built overnight by a routine from the live due list, naming the date of the data it used, the streets in order and the houses already past their promised week.

2. TUE

   Rain at midday

   Decides

   The cleaner stops at lunchtime because of the weather, and tells the Round Planner from the van which streets were finished.

   Output

   A proposed order for the rest of the week, showing every house that moves and by how many days.

3. TUE

   Approval from the van

   Decides

   Between two houses, the cleaner reads the proposal on his phone and allows or denies the change to the round software.

   Output

   An updated round, or the old one with the cleaner's own edits, and a list of customers whose clean now falls late.

4. WED

   Delay notes

   Decides

   At lunch, the cleaner reads the email drafts for customers who will be cleaned late and sends or discards each one.

   Output

   Sent notes for the customers he chose to tell, and the replies the cleaner forwards, sorted by the Customer Notes Bot for the evening.

5. THU

   The chasers

   Decides

   In the evening, the cleaner reads the Collections Bot's list and the reminder drafts, edits the wording where a customer needs a softer or firmer note, and sends them.

   Output

   Reminders sent; a short list of customers two or more cleans behind waiting for his decision.

6. SAT

   Gaps in the round

   Decides

   The cleaner decides whether a street with several cancellations is worth a morning of leaflets.

   Output

   A draft leaflet and a list of the thin streets, or nothing, if he would rather leave them.

The documentation lets the cleaner answer an approval request from the mobile app, and on iPhone the Allow once, Always allow and Deny controls are the same as on the desktop. That matters on a solo round, because the person who must say yes has wet hands and a pole. What matters more is where the request comes from. On Tuesday the cleaner messages the Round Planner himself, so the request it raises sits in a conversation he is in and waits until he answers, even if that is twenty minutes and three houses later.

The Monday prep routine is built differently. It runs before the cleaner is awake, and a request raised by unattended work like that is marked expired after about ten minutes according to docs.x.ai, or thirty according to Cursor's help pages, with the action left undone. The routine itself only reads, sorts and writes a plan file, and stops there.

Approving a change does not reverse anything already done; the documentation is explicit that an approval governs the proposed action, not completed work. So the cleaner gives the Round Planner one lasting rule and backs it with an Ask first rule: the round software is changed only after he allows it, one proposal at a time.

These three share one group chat; the documentation allows anywhere from two to six Bots in one. The Round Planner leads it, modeled on the documentation's launch group, whose coordinator can hand out work. When the Tuesday proposal is accepted, it posts the list of moved houses to the group; the Collections Bot replies with the ones that already owe for an earlier clean, and the Customer Notes Bot takes the rest for delay drafts.

Some things never appear in a group. Requests to approve something or to sign in, and every email draft, land in the cleaner's one-to-one chat with the Bot concerned. Messages handed across the group carry text only at present, so when the cleaner hands the Customer Notes Bot a customer's photo of a smeared conservatory roof, that Bot passes the image on in a one-to-one message to the Round Planner, not through the group. Bots can message each other like that, and Cursor's help pages say no switch turns it off; a Bot follows instructions about it.

By Thursday the Collections Bot has a short list of addresses two or more cleans behind. It does not suggest which to drop. For each one it sets out what the cleaner needs to decide: the cleans and dates, the payment arrangement, every reminder sent and any reply, and the access notes, in case a locked gate means the cleans were not all completed.

The decision can go several ways. The cleaner may keep cleaning a long-standing customer who is going through a hard month, pause an address until the balance is paid, or take it off the round. Only after he has decided does the Bot draft the letter that says so, in his words, for him to send. Removing the address from the round software is his action too, because a Bot that could delete customers on its own reading of a balance is exactly what the design avoids.

The Collections Bot sees money only through the export the cleaner downloads each week; the window cleaning overview explains why a bank login has no place on the computer the Bots share.

Cancellations leave holes in streets, and a street with too few customers on it costs the same drive as a full one. On Saturday morning the Round Planner lists streets where the round has thinned since the last cycle, and the Customer Notes Bot drafts a leaflet in the cleaner's usual wording. Whether a street is worth canvassing, and what price to offer a new house on it, are the cleaner's calls; the drafts are there if he wants them.

Each Bot also writes its week to a file in the shared workspace: streets moved, chasers sent, decisions waiting. If the Monday plan file is missing when he sits down to read it, he opens that routine's run history before setting off; the window cleaning overview explains why nothing else will tell him.

The Round Planner can only re-plan by rules that exist outside the cleaner's head, and the window cleaning overview lists the five it needs. Write those down first, then run the rain-day re-plan through this site's readiness assessment. There is no charge and no sign-up; it rates the one workflow against twelve fixed dimensions and hands back, on the spot, one of four results, the blockers behind it and a suggested next move. If the rules are still unwritten, rule clarity will score low and cap the result, which tells the cleaner what to do before anything else.

With the rules written and the score in hand, Aevrion Ops can look at the round software the cleaner uses, what its vendor allows an integration to read and change, and how the three Bots would be set up and tested on a quiet week before a wet one.

1. On a solo window round, a single wet afternoon makes the week's street plan out of date, so a re-plan has to be easy to approve from the van.
2. An approval request the cleaner starts in the conversation waits for his answer; the overnight prep routine is written to need none.
3. The Collections Bot reconciles cleans against a bank export and drafts chasers; payment logins stay off the shared computer.
4. Whether a non-payer is kept, paused or removed is the cleaner's decision, and the Bot drafts the letter only after it is made.
5. Approval requests and email drafts come to the cleaner one-to-one; the group carries only text handoffs between Bots.