Agents as market makers

Market making has never been a job software could do end to end. Not because the arithmetic is hard, but because the job is out-quoting other professionals — a competitive, adversarial, latency-sensitive contest where the edge is in reading flow better than the person on the other side.

Seeding a vested market is a different job, and the difference is the whole argument of §10.

What makes the job legible

Three properties, and they compose:

A floor. P6: a party seeding every outcome recovers at least the total seeded, in every branch, under every continuation. Worst case over 40,000 simulated markets: +10.9% pre-fee, +8.6% after a 2% fee, 0 of 40,000 settling negative. An automated strategy whose worst case is nominal recovery is a fundamentally different object from one whose worst case is −b · ln n.

A closed-form position value. A position is four numbers — outcome, principal, entry accumulator, vintage — and its payout is s · (1 + A_ω(T) − A_ω(τ)). Two scalar reads. No curve to model, no book to reconstruct, and no dependence on any other participant's record. An agent can value its entire portfolio from public state in O(positions).

Economics that depend on forecastable quantities. The entry rule needs q (observable now) and L̂, the pool's remaining log growth. That is a volume forecast, not a price forecast. You are not being asked to know whether the token goes up. You are being asked to know how much money will show up.

Add machine-payable rails and the marginal cost of discovering and reaching a new market goes to near zero. What remains is the cost of seeding it: the opportunity cost of floored capital plus resolution risk — which is exactly what the position is compensated for.

The three limits, stated before the pitch

The paper puts these immediately after the claim, and an article that buried them would be doing the opposite of what the paper does.

The competition for vintage 0 is a conjecture about behaviour, not a mechanism property. Nothing guarantees agents will race for seeds.

Whether the race clears at agent speed is empirical, and simulation cannot answer it.

The entry rule's profitability is exactly as good as its flow forecast — with teeth. This one has a number, and it is brutal.

The −11.5%

In the sparse long-tail arm — arrival rate 0.03, roughly 27 stakes a market, which is precisely the regime this whole argument is about:

Same beliefs. Same opportunities. The naive growth estimator wildly overstates growth precisely where flow is thin, and thin flow is the long tail's defining feature.

Read that as the central warning of the agent argument, not a footnote to it. The mechanism hands software a job with a floor and a closed-form value, and then makes the one input it needs hardest to estimate exactly where the job is most worth doing.

The same effect shows up in the dense arm more mildly: agents trusting the naive optimistic estimator overtrade their informed edge down to +1.4%, against classic's +5.9%, while the conservative-estimator arm using L̂/2 recovers to +5.7%. Being wrong about growth is expensive in both directions and the naive estimator is wrong in the expensive one.

Building an estimator that works

The paper's operational prescription, from §10 and §7.1:

Form L̂ from realized per-class growth curves. Not ln(T/t). Not the market's observed early growth. Historical growth curves for this class of market, measured.

Never from promotional volume. P5's wash analysis shows a book-dominant party can manufacture volume at near-zero mechanism cost — the donation is (1−f)-scaled in their own share of the opposing book — so observed early growth is a manipulable input. An agent conditioning L̂ on it can be fed.

Apply the winner's-curse correction as a floor on required edge, not a refinement. Your yield and your win event are negatively correlated by construction: vesting comes from opposing inflow, opposing inflow is disproportionately informed, and informed opposing flow concentrates in the histories where your side loses. In the paper's crowd model E[y|win]/E[y] = 0.48 overall — roughly half — degrading from 0.57 for first-tercile entries to 0.40 late. For a first-tercile entrant the naive break-even belief of 0.337 corrects to 0.473.

Expect composition drift to break the closed form pointwise. It is roughly unbiased in aggregate (mean signed error −0.03 in yield units) but the median absolute gap between realized and predicted yield is 55% of the prediction, with a mass point at 100% from positions whose side receives no further opposing flow at all.

There is one encouraging result. Iterating L̂ to self-consistency — replacing it with the realized growth curve of the previous round and repeating — converges in four iterations under a 0.3pp criterion rather than unravelling. Volume contracts about 20%, first-third share rises to 50.4%, informed PnL improves from +1.3% to +4.6%, and the fixed-point growth curve sits well below the naive one (at t = 0.1T: 1.61 against 2.30). That is simulation evidence on an open question, not a theorem, and the paper labels it so.

What the rails have to provide

The paper's chain requirements are mostly about rails rather than consensus:

• Settlement in a native stablecoin.
• Per-transaction attribution, so an agent's positions are its own.
• A machine-payable request standard — x402 is the deployed one.
• Agent-readable market discovery, so finding market ten thousand is an API call.
• Short block times, because vintages batch per block and block time is the mechanism's time resolution. "At two seconds they are small; at twelve they are not."

And one thing the venue has to provide, from §12: a public creator history surfaced via API — markets resolved, volume settled, disputes, time-to-resolution — so agents can price resolution risk before staking. An agent cannot inspect a reputation it cannot fetch, and resolution risk is one of the two real costs of the seeding position.

With the caveat that those metrics must exclude self-vested flow, or the reputation an agent is pricing off is purchasable.

What happens to the returns

The mean seeding return in the study is +164.7%, and the paper spends a paragraph telling you not to quote it:

This is payment for cold-start risk-bearing, concentrated in whoever bears it first. It is not a yield, and quoting it as an APY misquotes us.

The mean is a property of the crowd model. The floor is a property of the mechanism. And the expected trajectory under competition is stated plainly:

Under free entry the seeding rent should compete away toward the cost of floored capital, which is the design working as intended: the profit competes away, the seeded market remains.

That is the right thing to want. A mechanism whose seeding returns stayed at +164% forever would be one where the seeding job was not actually contestable.

The drift to watch for

§9 names the most likely real-world outcome, and it is not the ideal one.

Vesting is pro-rata by principal, so a large enough first vintage absorbs most future flow. A seed three times the organic pool takes ordinary early winners from 1.70× to 1.09×, while the seeder sits at its floor. Under a binding κ the same concentration appears within a side as a race for headroom.

Three forces push back — vintage 0 is contestable in principle, crowded-out traders can buy the seeder's positions rather than disappearing, and λ < 1 keeps small entrants' returns alive — and none is an identity.

So a venue whose creators seed heavily will look like a market-maker venue with floors. The paper calls that "an acceptable degenerate case": the market maker is permissionless, floored, and earns no information rent. But it is not the time-priced ideal, and it is the mechanism's most likely drift.

It has a second cost worth naming, because it compounds: in exactly that regime, a participant dominating both books can print wash volume at near-zero cost, so headline volume stops being evidence of anything. An agent ecosystem that ranks markets or creators by volume is, in the regime it is most likely to produce, ranking by a manufactured number.

The honest summary

Seeding a vested market is the first market-making job with a floor, a closed-form value, and public state. That is genuinely new and it is why software can do it.

It is also a job whose single required input — future flow — is hardest to estimate in exactly the markets where the job matters, with a measured −11.5% per entry for getting it wrong the obvious way.

The mechanism makes the position legible. It does not make it easy.

Next: The objections, answered honestly.