Two halves. First the agents' day — who wakes when, what each one reads, and what it is forbidden to touch. Then the four steps that get you there, with proof counted by your own team at each one.
Scheduled, capped, logged. Every run ends with a verdict a person can read in one line, and UNKNOWN is never OK.
What changed overnight, what the other agents flagged, and the three items that need a person today. Sent to two named people and nobody else.
You pick it at kickoff: a reconciliation, a chase list, a daily sheet somebody rebuilds by hand every morning. It runs before anyone arrives, with counts. They check a list instead of building it, and the hours returned are tallied by your team, not by us.
Uptime, spend per seat against its ceiling, credential expiry, unusual usage, and our own support-hours ledger. It warns at 80 percent and pauses at 100. It is the agent that protects both sides of the agreement.
Per account, per line: a robust z-score against a trailing window, so a number that is quietly sliding is caught while it is still small. Fifteen flags a week per owner at most, the rest to a watch list. It never contacts anyone. A person decides, every time.
Hours returned, flags raised and acted on, minutes saved per task, active seats, spend by line, cap events, build hours used. It never estimates a number it cannot count. One copy to you, one to our vault.
Four steps, in this order, with proof counted by your own team at each one. Nothing gets built that was not written down first, and nothing ships that a person has not read.
A paid discovery. We inventory your systems, name the source of truth for every fact, pick the one boring task to start with, and write each agent's spec with its cap before anything runs. You get a systems audit either way.
Two weeks · credited to the buildRead-only connections into what already exists, a knowledge base seeded with your rules, calendar and voice, and the first agent running in assist mode inside four weeks, with its proof line agreed before it starts.
Four weeks · one task, one ownerEach wave earns the next. Agents climb the trust ladder only against the accuracy threshold in the agreement, and the routed questions from each week become the agenda for the next.
Wave by wave · proof at eachOne flat monthly: uptime and the watchdog, model and security updates, the monthly proof report, a weekly check-in, and build hours for what you ask for next. Anything beyond them is a written change order with a fixed quote.
Flat monthly · never hourlyOne agent reads everything that changed overnight and says it out loud in under a minute. Items that need a person are named. Nothing in it was decided by the fleet.
This one is generic. Every industry page carries a brief written for that industry, in the words that industry uses:
That is the whole first conversation. We map what you run, pick the task your people would most like back, and show you what the first agent would do before it does anything.