How this actually works

Why does the agent dig at all? Because every session starts blank. An agent has no yesterday — and the codebase itself carries no knowledge about itself. There is no place where 'these 14 files depend on that function' is written down; that truth exists only while someone is deriving it. So the agent does the only thing possible: it reconstructs the world from raw text, brilliantly, every single time.

The cost isn't one dig — it's the multiplication. Every question, times every agent, times every day, forever. When eight agents answered this same question in one night, they spent about a million tokens producing eight identical copies of one fact. Nothing was wrong with the agents. The world they work in has no memory — that's the disease.

How this actually works

Each of these failures is mechanical, not stubbornness. A reminder is just text competing against an agent's training, and its training says: verify. A 'trust me' label fails because trust is built from repeated experience of reliability — and an agent with no memory has no experience; every session meets the map as a stranger. Even a receipt gets re-checked, because checking is what a careful engineer does with an unfamiliar claim.

The lesson took us five experiments to accept: any design that requires the agent to choose differently is doomed, because the choosing happens in a place we cannot reach. So stop reaching for it. Change what's beneath the choice.

How this actually works

Here's the concrete move. An agent touches the world through a handful of tools: the search it runs, the briefing that starts its task, the check that runs before it edits a file, the gate that judges its work. We run the agents' environment — so we rebuild those few tools, and their results simply arrive carrying more truth. The search returns matches plus the dependency map plus a warning about the one place text can't see. The briefing is assembled from the ledger. The pre-edit check injects the blast radius before the file even opens. The agent changes nothing: it reads its tool results, exactly as it always did.

Why is this transformational? Because it deletes adoption as a concept. Spell-check never asked writers to consult a dictionary — it put the dictionary under the keystroke. Maps never persuaded anyone to plan routes — the route arrives inside the act of asking. Every technology that truly won disappeared into a motion people already made. This converts an unsolvable psychology problem — 'make agents trust' — into a solvable infrastructure problem: 'make the ground truthful.' Psychology has no fix. Plumbing does.

How this actually works

Mechanically, a deposit is small: when work derives a truth, a hook writes it down — the claim itself, fingerprints of the files it depends on, a re-runnable command that proves it, and the commit it was true at. Reading it back is even smaller: check whether any fingerprinted file changed. Unchanged — the fact holds, use it. Changed — the fact expires by itself, and honesty costs nothing.

This inverts the economics of knowledge. Today, a nine-minute derivation is an expense that evaporates at session end. Deposited, it becomes an asset: the first use costs nine minutes, every later use costs a second — and each new fact makes the next task cheaper. It's the difference between a workforce that burns effort and one that accumulates it. Agents stop being an oral culture, where every generation re-derives everything, and get what writing gave us: the ability to build on what's already known.

How this actually works

Why believe the ledger? Not because it promises. Because each fact carries three things that make lying structurally hard: a receipt — a command anyone, human or agent, can re-run to reproduce the claim; a currency stamp — the fact auto-expires the moment its basis changes, so stale truth can't masquerade as fresh; and a fence — an explicit list of what this fact cannot see, so silence is never mistaken for 'all clear.'

And above the facts stands a standing adversary: an independently-built recount that recomputes reality from scratch and diffs it against the ledger's answers, on every change, forever. That system caught the map under-reporting by 63 files on its very first day — before any decision was built on the error. Trust here isn't a feeling. It's a record: attacked continuously, and it survived — and you can run the attack yourself.

How this actually works

Compounding, practically: today the ledger holds one fact, so the gate pays full price — 124 checks from scratch. But every gate run, every briefing, every closing session banks more. As deposits cover more ground, the same gate re-verifies only what actually changed and skips the rest with proof in hand. The factory literally gets faster every week it runs on unchanged ground — infrastructure with increasing returns, where every other factory on earth runs at constant speed forever.

And the contract ports. The same four-part fact — claim, receipt, freshness, fence — works for 'what services does this veterinary practice offer' and 'what did we decide, and why.' One core, many domains: code first, business truth next, intent last. That's why the scoreboard is the only proof we accept: infrastructure stories are cheap to tell, but counters can't be argued with. When facts minted, reads, and re-derivations avoided climb week over week, the idea is real. Until then, it's a bet — a precise one.