Earned to the cent

A few nights ago a friend in the backyard asked what I actually do now. Instead of explaining, I pulled up a live session on my phone. Not a demo, a real project running right then, on a system that thousands of loans depend on.

Here is what they were looking at.

The kind of project that eats teams

A company runs a loan servicing system on a decades-old desktop application. It is the system of record. Real customers, real money, every day. They had already paid for one modernization that never reached production and is now being thrown away. This is the second attempt at replacing something that cannot go down and cannot be wrong. The bar is brutal: the new platform has to be provably correct against the old system's math before anyone is allowed to trust it.

The twist is who is doing the work. It is not people writing code the usual way. It is an agentic team of four AI personas working under a constitution: gates that block, reviews that return cited conditions, math that proves itself. One runs product and discovery. One runs the build under strict test-first discipline, one test at a time. One keeps improving the rulebook itself. One plays the adversary whose only job is to ask "is this actually true?" and check every claim.

Why this is not vibes

I need to be precise here, because "AI wrote the code" usually means "we hoped for the best." This is the opposite.

Money is exact decimals, never floating-point. A rounding error at a loan servicer is a real-money defect, not a rounding convention. The calculation engine is proven correct to the cent against the very legacy system it is replacing, before any behavior is allowed to change. We never re-derive the old math from scratch and hope.

And the tests have to earn their keep. The system actively tries to break the code in small ways and checks whether the tests catch it. If they do not, the tests are strengthened until they do, before anything ships. So you cannot quietly weaken a test to get to green, and you cannot ship without a review panel that returns specific, cited conditions you have to meet first.

The human in the loop is the client, not me

This is the part I care about most. The work is validated continuously by the people who own the system, not handed over a wall at the end. Every slice ships to a site where their own tester runs the real thing. The specification travels with the code, in the same place, so the two can never quietly drift apart.

Last week their tester found something: two interest-schedule conventions were diverging by small amounts. That one finding kicked off a governed study against the legacy engine, a root-caused fix, and a set of tests that lock her exact scenarios in permanently, to the cent for both conventions, so the same divergence can never recur silently in a future change. She found it. The agentic team characterized and fixed it under governance. It is guarded for good.

I am the orchestrator, not the coder

I am not writing the implementation. I am not writing the tests. I work at the level of intent and judgment: build this slice next, prioritize her fix and then ship, make tenant isolation visible from day one, keep the specs in sync as we go. I relay the client's feedback, make the product and business calls, and decide sequencing. The team turns that intent into governed, test-driven, proven software, and it pushes back with cited conditions when something is not sound.

The constitution is what makes that safe. Nothing stays hidden, because the agents have to surface their shortcuts and their open debts. Everything stays aligned, because every artifact traces back to a rule. And everything stays auditable, because there is a receipt on every gate and every commit. That is what lets me orchestrate instead of micromanage.

Earned, not asserted

That is the whole point. A modernization that failed the first time is being rebuilt by a governed agentic team, with the client as the continuous check and a human giving intent. It is not finished, and I would not claim it is: some pieces are live, more are in progress, and each one earns its trust the same way. That trust is not a claim in a pitch deck. It is earned, to the cent, against the system it is replacing.

Back in the backyard, my friend asked the obvious next question: so what do you actually do? The honest answer is that I point a capable team at a hard problem and hold the standard. Everything else earns its trust in the open.