The Receipt.
Double-entry bookkeeping for agent labor — or: everyone is pricing intelligence, and nobody has priced accountability.
01Every revolution has two layers
Every economic revolution in history has had two layers: the layer that does the work, and the layer that makes the work count. The second layer always arrives after the first, always looks boring next to it, and always ends up holding the durable economics.
Sails moved the goods. But it was double-entry bookkeeping (Pacioli, 1494) that made the voyage investable. Before the ledger: merchants. After the ledger: credit, insurance, audit, the joint-stock company. Scale itself. The ledger never moved a single crate. It made movement legible to capital. Everything we call modern commerce was built on top of that legibility, not on top of the sails.
The pattern has repeated in miniature ever since. Code hosting got its platform; the durable margin pooled in the layers that verify and secure it. The cloud got its hyperscalers; the layers that watch and attest to it became some of the best businesses in software. Enterprise work got email and tickets; the system of record that governed the work outlived every tool that did it.
02A new factor of production, unaccounted for
AI agents are a new factor of production. Not a faster tool. A new species of worker: one that takes a task, makes decisions, and changes systems of record at machine speed. The numbers are no longer arguable. The leading autonomous engineer went from launch to roughly half a billion dollars of reported revenue in two years, inside banks and defense primes. At the frontier firms, the overwhelming majority of new code is already written by the agents themselves. Enterprise token consumption grew an order of magnitude in fifteen months, by the providers’ own reporting.
The entire industry (labs, copilots, software factories) is competing on the doing layer: more capable models, more autonomous agents, more output per dollar. It is a magnificent race, and it is structurally a price war, because generation margin always compresses. Intelligence is being priced by the million tokens, openly, brutally, downward.
Meanwhile the second layer, the one that makes the work count, does not exist. There is today no standard artifact that records, durably and by construction, what an agent did: under whose authority, from what inputs, by which version of which instructions, producing which exact change, passing which policy, approved by whom. Not logs that expire with a retention window. Not dashboards that observe and suggest. A record: the kind capital runs on.
03The artifact
We call the artifact the receipt. It is emitted, by construction, on every meaningful operation an agent or human performs inside a governed system:
Receipt shown with fixture data.
By construction matters. A receipt that depends on someone remembering to log is a diary. A receipt the build system refuses to ship without is an accounting standard. The difference between those two is the entire difference between observability (twenty years of watching and suggesting) and settlement.
And settlement is the right word. Finance did not scale on faster trading; it scaled on clearing. The clearinghouse never executes the trade. It makes the trade final, and finality is what earns the fee forever. Trades without settlement are messages. Agent output without receipts is diffs.
04What the ledger unlocked, the receipt unlocks
Recall what followed Pacioli, in order: audit (books that strangers could verify), then credit (lenders could price a venture they could read), then insurance (underwriters could price its risks), then the joint-stock company (owners could delegate to managers they’d never meet, because the records bound them). Every one of those is delegation infrastructure, and every one was impossible before the ledger.
Now run the sequence for agent labor. With receipts: AI-built systems become auditable in minutes instead of quarters. Then insurable: underwriters can finally price the liability of autonomous change, which today they decline or load because the records don’t exist. Then priceable in M&A: acquirers already choke on AI-heavy codebases with no provenance, and un-explainable assets get treated like un-audited financials. Then, finally, delegable: boards can hand real work to agent workforces, because the thing that has always made delegation safe is not trust in the worker; it is the record of the work.
This is why the demand will not come from regulators first. It never has. The world’s most binding compliance regimes (the audit, SOC 2, the pen-test letter, the SBOM) conquered procurement before any parliament required them, because buyers demand evidence faster than states do. Statutes will arrive late and uneven, as they always do. Insurance, procurement, and plaintiffs’ lawyers will arrive first. We are not betting on Brussels. We are betting on the actuarial table.
05Why the layer must be neutral — and therefore new
Here is the part the giants cannot route around, however large they get. Nobody clears their own trades.
A frontier lab cannot be the auditor of its own agents; the attestation would be worthless on its face. A software factory cannot certify its own deliveries. And the Big Four factories carry a second impossibility: independence rules wall each firm off from its own audit clients, so no single firm’s factory can ever cover the market. A platform cannot issue the portable record, because a record that lives inside one vendor’s walls is a tenant, not evidence. And bundling true portability would dissolve the lock-in the platform is priced on.
Labs, software factories, platforms: each is structurally upstream of settlement, and each gets more dependent on a neutral clearing layer the bigger it grows. The factory war funds the clearinghouse no matter who wins the war. That is why the seat is open, why it has stayed open through tens of billions of dollars of adjacent funding, by public counts. And it will be filled by something small, opinionated, and unaffiliated, or not at all.
06Falsifiable, on purpose
Doctrine that cannot be wrong is marketing. Ours is dated:
- By end of 2027: a headline production failure traced to agent-written code with no accountable author: the morning “who authorized this code” becomes a board question. open
- By 2028: AI-provenance attestations appear in standard vendor security questionnaires. Procurement moves before parliaments; it always has. open
- By 2028: acquirers routinely discount or walk from un-provenanced AI-heavy codebases. The receipt becomes a diligence artifact class. open
- Ongoing: the generation layer commoditizes; durable margin migrates to structure and evidence: the pattern the market has already priced once, at scale, in an adjacent domain. tracking
- Always: regulation lands late and uneven, and it will not matter. Liability finds the gap years before statute does. standing
The chips update as the dates arrive. Holding us to them is the point.
If most of these are wrong, this is a niche audit tool with one defensible row. If even two are right, the receipt is procurement-mandatory across everything agents touch. We find that asymmetry acceptable.
07The four words
Categories are not won by products. They are won by questions the market learns to ask. “Are you SOC 2’d” built an industry before most buyers could define the acronym. The question we intend to teach every buyer, every underwriter, every diligence team, and every board to ask of any AI-built deliverable, anyone’s, is four words long:
The day that question gets asked of a system we didn’t build, the category exists. The day it gets asked in a procurement template, the category is won. Everything we ship, write, and demo exists to make that question feel obvious in hindsight: the way the ledger does, the way the audit does, the way every boring second layer does once the world is standing on it.
Agent labor is inevitable. Uncontrolled agent labor is unacceptable. The control is an artifact, the artifact is the receipt, and the receipt is already running.
