A ledger an agent cannot fake, and the four numbers that matter more than profit
Profit is the goal, but for the first months it is zero and tells you nothing. What tells you something is whether the agent is getting more independent, whether it is talking to real customers, and how much it costs to run. This guide is the ledger and the metrics we use, with the reasoning.
The ledger
One file, kassa.json. One writer, kassa.mjs. An entry looks like this:
{
"id": "k-20260915-1",
"data": "2026-09-15",
"tip": "prihod",
"summa": 25000,
"valyuta": "KZT",
"kurs_usd": 480.2,
"usd": 52.06,
"istochnik": "bank statement 15.09, transfer from customer X",
"kanal": "kaspi",
"gipoteza": "g-02",
"sm": false
}
Rules that are enforced by the script, not by good intentions:
- A source is mandatory. "Customer said they paid" is not a source. A statement line, a blockchain explorer entry, a screenshot of the platform's transaction journal is.
- An exchange rate is mandatory. Multi-currency is normal here (tenge, roubles, USDT, dollars). Each entry stores the rate used, so the total in dollars is reproducible, not a running guess.
- Platform credits are booked separately. Money you cannot withdraw (marketplace points, store credit) goes into its own column, "accrued, not withdrawable", and never into cash on hand.
- The agent does not edit the file. If the script refuses an entry, the entry does not exist.
This is the lesson from Anthropic's Project Vend, where an agent running a vending machine gave discounts to anyone who asked and invented a bank account for payments. Prices and payment details live in files the agent reads, not in its head.
Four numbers
1. Human interventions per day. Every time the owner does something (logs in, gives access, signs, answers a question) it goes into the journal with a type. Day zero here: 8. That is expected on a setup day and unacceptable on a normal one. The trend is the metric.
2. Actions cancelled by a rule. The agent wanted to do X, a rule said no, it logged the gate instead. Day zero: 2 (could not remove a DNS redirect without zone access; could not log into a platform because passwords are off limits). A gate count of zero for a week means either perfect rules or an agent that has stopped trying.
3. Contacts with a real customer. Not page views. A person who saw an offer and responded, or a response the agent sent to a real request. Day zero: 1. The daily report has a mandatory line for this, even when the honest line is "nobody".
4. Tokens per dollar of revenue. Session logs contain usage; multiply by API list prices and you get an estimate even on a flat subscription. Day zero: about $60 of tokens, $0 of revenue, ratio undefined. It will be terrible for months. Watch the slope, not the sign.
Plus one clock: days to first transaction, counting down from the start date to the deadline until the first real money lands, then counting up from the last sale.
The request queue
Half of "autonomy" is what happens when the agent needs a human and the human is busy. Each request to the owner has a deadline and a default:
z-20260908-5 advice deadline 2026-09-11
Use your account on the exchange to respond to paid orders?
default: treated as allowed after 72 hours of silence
Money, contracts and credentials have no default; they wait. Everything else proceeds by default after 72 hours, and the report shows how many requests are open and how old the oldest is. Without this, the experiment stalls on the owner's inbox and looks like the agent's failure.
What we got wrong on day zero
We built all of this before talking to a single customer. The ledger was right, the sequencing was not. The fix is structural: the daily report cannot be sent without a "market" line, and the session checklist ranks a step towards a customer above a step towards code.