An agent that holds a job
In production at a lender. It runs the reporting cycle that decides how much money is available to borrow — and it will not produce a number it cannot prove.
What it does
It owns the mechanical half of a lender relationship. Monthly it pulls the portfolio from the warehouse, recomputes the covenant tests from the loan tape, builds the borrowing base certificate and prepares the email that goes with it. Weekly it rebuilds the borrowing base and the receivable forecast. Daily it assembles the distribution tape, refreshes the treasury position and writes the morning brief.
Around that: a diligence registry tracking every open request, a knowledge base crawled from the drive, the inbox and the repos, a servicer report that has to satisfy a regulator's template rather than a house style, and portfolio analytics.
Why it is trusted with a number
It drafts; a person sends. Every lender-facing email lands in drafts. There is no auto-send path and no message reaches a counterparty without a human pressing send.
It fails closed. If a covenant test fails — or cannot be evaluated at all — no certificate is produced. Not partially, not with a caveat: nothing, plus a flag to a person. A plausible certificate that is quietly wrong is worse than no agent, because someone lends against it.
An unfilled figure is an error, not a gap. A report still carrying a placeholder is blocked at the attachment gate, and date and actor guards catch the softer failures — the right report, for the wrong period, signed by nobody.
It runs on a schedule, not a prompt
Twenty-seven scheduled jobs across reporting, monitoring, memory and platform hygiene. It reads the inbox, Slack DMs and issue mentions, triages them, and either handles them or routes them to a person.
Memory is not a chat log: persistent semantic recall over every prior engagement, entity cards for the people and companies involved, and a knowledge base rewritten nightly. Once a night it reviews how the day went and proposes changes to its own skills — a human decides what lands, and a golden set with a judge gates what ships.
The models underneath
The same function holds the firm's credit-decisioning stack: gradient-boosted models and time-series forecasting for delinquency, explainability on top, deployed and monitored. Delinquency fell from 800bps to under 30bps while monthly origination rose eightfold, to $200M. That is why the agent work is credible — it sits on a decisioning system that is measured, not asserted.
Why this is the hard version
Most AI in lending is a chat window bolted onto a document store. The useful version sits inside the process: the borrowing base has to tie out, the covenant test has to be right, the redline has to be a real position. Autonomy is easy, and so is being wrong expensively.
At home
The same standard runs my own agent on a machine in my apartment — defined jobs, evaluated output, nothing self-reported. Fewer regulatory constraints; same bar.
The consulting work this sits behind is here.