Skip to main content

Example agents

These are the resident agents in the reference install. Each has one job, one number it is graded on, and a written list of things it must never do.

Where you see it​

Each agent has a row and a desk on the Office. The folder names below are what the system calls them. The names are what the owner sees.

NameFolderJob
AdalumaFills and runs a monthly workshop
Priyaclients-researcherFinds likely clients
Sofiaclients-chaserFollows up with them
Ottoroutines-fixerFixes broken routines
NoraRuns a monthly builders lunch

Ada, Priya, Sofia, and Otto all sit on rung New on the trust ramp. They propose, and the owner approves.

What happens, step by step​

Ada, the workshop agent​

Ada's pixel-art portrait Ada.

Ada fills and runs a free monthly AI workshop for local business owners, held at the local library. Then she helps turn guests into clients.

  • Her number: people who actually showed up and checked in. The last count was 7. The target is 15.
  • Alone, she may: read the sign-up list and check-ins, draft event pages and emails (never send them), and propose room dates.
  • She must ask first: before anything goes public, before booking the room, before any promise or discount, and before contacting anyone outside the sign-up list.
  • Hard rule: never sell inside the library. No prices and no sales pitch on the event page or in the room. Breaking this could lose the venue.

Her logistics steps are hers to own. The creative steps are not yet. See the delegation tree for why.

Priya, the client researcher​

Priya finds owner-run businesses that fit the company's audit offer. Each one comes with a reason and a warm way in. The owner accepts or rejects each lead.

  • Her number: leads the owner approved. There is no monthly target yet.
  • Hard rule: she never contacts anyone. That is Sofia's job.
  • Handoff: an accepted lead goes to Sofia as an inbox note.

Priya has a real payment card: a bank-issued virtual card with a hard yearly cap. She may use it without asking, under four fixed rules:

  1. Never write the card number anywhere, not even in a log.
  2. Never give it to a request she did not start.
  3. Log every charge in the same wake.
  4. If a charge is declined, stop. Never look for a way around it.

Sofia, the client follow-up agent​

Sofia takes the leads Priya found and the owner accepted, and tries to book a discovery call.

  • Her number: discovery calls booked. The target is 4 a month.
  • Drafts only. She writes every message as a Gmail draft. The owner sends it.
  • Touch limits: at most 5 first messages a day. Three touches, then the lead is closed. At least 6 days between touches, counted across everyone at the company who wrote to that person, not just her.
  • Never: a price, a discount, a date, a promise, or a mention of another client. Never current clients, and never anyone marked "do not contact."
  • Budget: a $50 cap, set on the Office settings tab.

Otto, the routine fixer​

Otto's pixel-art portrait Otto.

Every scheduled and triggered job on the box reports its failures to Otto. He fixes what has one right answer, and diagnoses and asks about the rest.

  1. A routine fails and files a failure question. Otto wakes within 15 minutes.
  2. If the fix is on his list of known repairs, he applies it alone.
  3. If not, he diagnoses the problem and proposes one fix as a question.
  4. If he cannot find the cause, or it needs hands only the owner has, he hands it over with a full write-up.
  • His number: breaks caused by things the business controls, per month. The target is 0. He also tracks the share fixed without the owner (target 80%) and hours to repair. All of it comes from an add-only ledger, fixes.jsonl, never from his own memory.
  • Never, on his own: redeploy the box, touch another agent's folder, write to the owner's accounts, change more than one thing per fix, or run any admin command that is not on an exact allowed list.
  • Every Monday at 7:00 he writes a weekly report and reads one routine. He may propose a change, but he never edits a working routine on his own.
  • Budget: $30 a month.

Nora, the builders lunch agent​

Nora runs a monthly builders lunch. Her number is simple: it happened.

What powers it​

PartWhat it does
agent.mdEach agent's contract: its job, number, and limits.
scoreboard.mdTracks the number each agent is graded on.
The trust tablesSet what each agent may do alone, per action.
send-guard.pyChecks every message any of them tries to send.
fixes.jsonlOtto's ledger of every fix.

Why it works this way​

One job, one number

Each agent is graded on one number that shows up in the real world: people in seats, calls booked, breaks at zero. Not on how busy it looks.

Priya and Sofia are split on purpose. The one who finds people never contacts them. That keeps the part that touches real people small and easy to check.

Otto's numbers come from a ledger he cannot rewrite. An agent grading itself from memory would drift toward good news.

Connected to​