What Agents Can't Manufacture

The cold open

Here is this week’s thought experiment, posed exactly as it ran:

In an agentic market, what becomes the most valuable asset? The last decade rewarded the best data. This decade looks like it rewards the most compute. But if agents become the primary buyers, sellers, negotiators, and optimizers — what becomes truly scarce?

It reads like a ranking problem: line up data, compute, distribution, trust, and crown the heaviest. That framing is the trap. The question is not which moat is biggest — it is which moat the agentic cycle is busy filling in. When a question asks you to name the most valuable asset, check first whether you are naming the one currently going abundant.

The vote

The poll closed with 52 votes. Here is how the room split:

The reflex answer is Data — the asset that won the last war. That instinct is not wrong about the past; it is wrong about the tense. I voted Trust against that reflex, and the room closed on the same side: a 52% majority, Data second at 33%, Compute and Distribution tied at 8% each. The margin is final now, but the reasoning underneath it is the same, and the rest of this is why the column I picked was the right one.

The reframe

Start with a pattern, not a law. Scarcity tends to relocate, not vanish. Some constraints really do dissolve for good — clean water, literacy, long-distance reach all got durably un-scarce. But inside a single technology cycle the binding constraint usually moves rather than disappears, and it moves using the exact capability that made the previous moat cheap. Data was scarce until synthesis and embeddings turned any corpus into something interchangeable. Compute was scarce until per-token prices collapsed into a rental commodity off someone else’s depreciation curve. Distribution was scarce until agents began querying sources directly and routed around the channel.

Notice the pattern: synthetic-data abundance is provenance scarcity; cheap-optimization abundance is intent scarcity; channel-bypass abundance is accountability scarcity. The crowd names the asset that just became reproducible. So consensus on the moat is a lagging indicator — it points one cycle behind. You predict the next scarce thing by naming the residue of the current abundance.

The Self-Supply Test

Now kill the three losers with one instrument. An asset is durably scarce in an agentic market only if the agent cannot acquire it by spending its own cycles or its own money. Call it the Self-Supply Test, and run all three rivals through it at once.

Data is synthesizable — a smarter agent manufactures more of it, cheaper, on demand. Compute is rentable and fungible — the same FLOPs sold to your competitor by Tuesday, off the same depreciation curve, leaving no buyer a durable edge. Distribution is routable agent-to-source once every buyer can address every seller directly. Every one of them gets cheaper as the agent gets smarter. Their abundance is produced by the precise capability the poll is asking about. An edge that the rising tide of intelligence erodes is not a moat; it is a sandbar.

Object that agent platforms will re-concentrate distribution — that the model providers become the new channel — and look at what that concentration actually is: a permission position, the right to route and to be the default. That is not a data moat or a compute moat. It is the trust moat wearing a logo.

Why trust survives it

So why does trust survive the same test? Because of where its value comes from. An asset stays scarce only if its value is conferred by a counterparty who would lose by faking it. Trust is not a property of the trusted party; it is a standing granted by someone with something at stake.

This is the structural reason an agent can never self-supply it. A self-signed certificate is worthless. A self-issued credit rating is a joke. The signer cannot occupy the position that produces credibility, because that position is defined by being external to the signer and exposed to loss if the signer is wrong. Staking looks like the loophole — an agent posting its own capital to be believed — but it isn’t one: the collateral only counts if it is denominated in something a counterparty already honors, and it only bites if an outside party is positioned to rule you wrong. Bonded trust still bottoms out in external standing; the agent rents the judge, it never becomes one. No quantity of intelligence closes the gap, because the missing input is not computation — it is external standing. It can fake every input; it cannot fake the one thing that, by construction, must be granted from outside it.

The Permission Bottleneck

Ground it where you and I actually work. Put an agent on the buy side, an agent on the sell side, and let the negotiation be two models reconciling bids. Every coordinate either side cares about — the audience claim, the performance number, the brand-safety signal, the bid itself — can be synthetically generated by whoever benefits from it. When every input is self-mintable currency, every input collapses to zero differentiation. The only non-fabricable coordinate left is whether the counterparty’s claim is true.

And here is where the scarcity actually binds: not at inference, but at execution. An agent’s value is gated on whether its output is permitted to act unreviewed — the Permission Bottleneck. A brilliant bid the agent isn’t trusted to place is worth nothing; a mediocre one it is trusted to place captures the whole transaction. Trust is the single input that decides whether intelligence is allowed to touch money. Unlike data it isn’t copyable, unlike compute it isn’t rentable — it is slow to earn, non-transferable, and forfeited in one bad print.

The Opacity Penalty

Now the turn that should sting whoever voted Data. The contrarian’s lazy version of this argument is “stockpile trust like an asset” — and that is its own category error. Trust is not a vault you hoard; it is a credit line a counterparty extends and reprices on every request. Operationalized, it reduces to two things: how cheap you are to verify, and how much you’ve posted that gets slashed when you’re wrong. An attestation with no downside is just a more confident assertion, and adversarial agents discount it to zero.

Which inverts the last decade’s prize. From a counterparty agent’s view, a large private data lake is an unauditable black box — the Opacity Penalty. Its claims can’t be checked against an external witness, so a rational agent discounts them or demands expensive proof. The party with less data but full, signed, externally reconstructable provenance gets believed faster and cheaper. Scale of private data increases opacity; opacity increases verification cost; and verification cost is exactly the thing every agent is minimizing. The asset that won the last decade becomes a tax on being believed in this one.

The write-in votes

The poll offered four doors. The strongest answers refused all of them and wrote in the margin — and each write-in, run through the same test, lands back on the same column.

“Energy.” The most physical answer on the board, and the most honest about what compute actually costs to run. But energy is scarce the way oil is scarce: metered, fungible, purchasable. You close an energy gap with a checkbook — it is a commodity constraint, not a conferred one, and commodity constraints are exactly what the cycle dissolves. Energy gates how much an agent can think. It is silent on whether anyone lets it act.

“You need to trust the data.” This arrived as a vote for data and is the cleanest statement of the thesis on the page. Read it again — not “data,” but “trust the data.” The sentence ranks them for you: data is the object, trust is the predicate. A lake you can’t trust is a liability you pay to store; a handful of records you can is a credential. When the best case for data has to borrow the word trust to finish its own sentence, the question is already answered.

“Capability is the only thing that matters — everything else rolls up into it.” The sharpest objection, and worth the most room. It is right about the rollup: capability does subsume data, compute, and distribution — which is precisely why those three fail the Self-Supply Test. But capability is the one variable every agent is racing to raise in lockstep, and a capability that rises everywhere at once is a baseline, not a moat. The move it can’t make is the last one: capability is something you build; permission is something you’re granted. The most capable agent in the market still cannot place the bid until a counterparty decides to let it. Capability gets you to the table. Trust is the seat. Roll the whole stack up into capability and you have still only built a better agent — you have not answered whose agent gets to act.

Four Fridays, one reservoir

Four weeks, four questions — standards, fraud, the recommendation-decision line, and now scarcity itself. Strip the costumes and, in hindsight, the four keep pointing at the same place. Standards migrated from a data layer to a trust layer. The fraud blind spot was never authenticity; it was accountability. The recommendation-to-decision line was an accountability handoff. Three weeks I kept finding the place trust was leaking out of. This week names the reservoir it pools into. The handoff was the symptom; trust is the thing being routed toward.

My vote, corrected

So let me be precise about the Trust ballot, because contrarianism for its own sake is just vanity. I did not vote Trust out of faith in virtue. I voted it out of arithmetic. Data and Compute are answers to “what makes a better agent” — a question this cycle is busy closing. Trust is the answer to “whose agent gets to act” — the only question it leaves open. Pick the asset whose scarcity the cycle deepens, not the one it dissolves.

Because every agentic chain, however deep and however smart, eventually bottoms out the same way. An agent can spin up a thousand sub-agents in a second, but it cannot mint a single one of their reputations — those are earned from the outside, slowly, by someone with something to lose. The chain terminates in a human, a brand, or an institution whose standing it can rent but never make — a reputation borrowed from outside the machine because the machine has no way to manufacture it from within. That terminus is the scarce asset. Everything upstream of it is becoming free.

So stop ranking the moats by size. Ask which one your agent can build for itself by getting smarter — and put your chips on the only one it can’t.

Where this goes

If trust is the one asset agents must borrow from outside themselves, the next question is who’s allowed to lend it: does the credential that lets an agent act come from a protocol it speaks, a constitution it’s bound by, or a contract it can be sued under? Whoever owns that issuance owns the bottleneck. The thread has since taken its sharpest turn: in the context economy, the permission that gates whether an agent may act turns out to be a relevance signal in its own right — consent, provenance, and scope tell the system which moment matters, which makes governance the engine of relevance rather than the brake on it. The Fridays since have kept pulling on exactly that thread — the running experiment lives at /now if you want to follow it live. My 2c, as always — food for thought for the weekend.