The First Agentic Navigator
Evidence essay · Signed commentary, not a badged data claim
This one is a personal note, not a data claim, so I am signing it. Every other essay on this site stands on a source registry; this one stands on thirty years of watching what institutions do when a fact is too expensive to check — and on the two years since that changed.
For most of my career, the binding constraint on economic analysis was never theory. It was reading. A country desk officer, a debt-crisis mission, a wealth-accounting exercise like the one I ran in 1995 — every one of them ran into the same wall: there is a mountain of qualitative record behind any number worth trusting, and no institution can afford to read all of it, so it samples. It picks the capital cities, the flagship indicators, the countries with the best statistical offices, and it extrapolates from there. The extrapolation is not dishonesty. It is what you do when reading is the scarce resource.
Frontier models changed which resource is scarce. Claude and its peers did not make institutions smarter about economics — the theory hasn’t moved. What moved is the cost of reading everything: every administrative filing, every footnote, every country’s own accounting convention, checked against the next country’s, at a marginal cost close to zero and with a session log that shows exactly what was checked and how. That is a different kind of change than a faster spreadsheet. It removes the reason to sample.
FAND is the claim, applied
The name is not a coincidence. FAND — the First Agentic Navigator for Development — is what this site is built on, and it is also the argument: agentic AI applied directly to the machinery of national accounts, not as a chatbot bolted onto the old workflow but as the research partner that makes a wider ledger affordable in the first place. The scale of it — 200+ of the 237 economies and 3,258 US counties, on one closed identity — is not a productivity anecdote. It is what changes when a single economist can direct an agent through hundreds of logged, checkable sessions instead of managing a department that can only afford to look closely at a fraction of the world.
| Sampling era (my 1995 capstone) | Full-read era (FAND, 2025–) | |
|---|---|---|
| Coverage | Select flagship countries and indicators | 200+ of 237 economies, 3,258 US counties |
| Why the gap existed | Reading and reconciling records was a department’s worth of labor | Marginal cost of one more county is an agent-session, not a hire |
| What backs a number | Institutional memory, expert judgment | A source registry and a reproducible, re-runnable pipeline |
| What happens to the footnote no one had time to check | It stays a footnote | It becomes a line, or a disclosed gap |
What this changes for the reader, not just the builder
If you sit in a governance or investment seat, the practical shift is this: “we didn’t have the analyst-hours to check that” stops being an acceptable answer, because the analyst-hours it used to cost no longer gate the check. Audit and replication — historically the parts of institutional analysis that got cut first when budgets tightened — get cheaper faster than the analysis itself does, because verifying a claim an agent already traced is easier than generating the claim was in the first place. That inverts an old institutional habit of trusting the summary and skipping the trail.
None of this is a claim that the model does the judgment. It doesn’t decide what a substantive right is worth, or which uncertainty band is honest — that is still the economist’s call, made and re-made across the session log this site keeps. What changed is only the reading. But reading was the wall. I spent thirty years on the other side of it, and I did not expect to see it move in my lifetime.
— John C. O’Connor