Yesterday I wrote about standing up the governance before any product code. Today the first real code exists — and I built it as a single, thin slice that runs all the way through the system, not as a pile of half-finished parts.
The temptation with a multi-agent system is to build breadth: a few connectors, a few analyzers, a report skeleton, all at 60%. It feels productive. It teaches you almost nothing, because nothing runs end to end.
So instead: one path, complete. Read some real data in (read-only, least privilege, fully logged), store it, have a single agent reason over it, produce findings, render one section of output. That's it. One narrow vertical, working, tested, demoable on a real repository.
The lesson that keeps proving itself: a thin slice that runs end to end tells you more in a day than four disconnected pieces tell you in a week. It surfaces the integration problems — the ones that actually kill projects — immediately, while they're cheap.
The part I'm proudest of is the invariant I mentioned yesterday, now enforced for real: every finding must cite the evidence it came from. Not "should." Can't not. I modelled it so that a finding without a source is impossible to construct — the type system rejects it before it ever reaches a database. There's a test whose entire job is to try to create an unsourced claim and assert that it fails.
In a trust business, "the system said so" is not a citation. If I can't click from a conclusion back to the exact artifact that supports it, the conclusion is a bug — so I made unsourced conclusions unbuildable rather than merely discouraged.
Breadth comes next. But breadth on top of a proven spine, not instead of one.
More as it ships. I'm logging the whole build in my Lab.