Market microstructure · Personal agents

The Intent Order Book

Binary bookings are lossy compression of intent. When personal agents hold graded, durable demand, markets invert: aggregate first, synthesise inventory second.

Scott Farrell · LeverageAI · ~18 min read

Ten days before Sunday, the inventory opens. You do not know what Saturday night will look like. You know only that you generally like Sunday pickleball, that seats fill, and that if you wait until your intent becomes firm, the liquidity will already be allocated to people who clicked earlier.

So you click BOOKED. The system records a fact. What you actually held was a probability distribution with conditions attached. The booking software flattened it into a bit.

That flattening is not a UI inconvenience. It is a market-design failure — and once personal agents can hold graded, durable intent, it becomes an unnecessary one.

The premature-commitment tax

A premature-commitment tax is what a market charges when it forces uncertain humans to crystallise fuzzy future intent into a hard state before they know enough to mean it. You pay it in attention (weekly polling), optionality (booking extras “just in case”), and truth (the provider’s fill board lies a little because soft demand is labelled firm).

I’m booking ten days in advance because that’s when the inventory opens. I don’t even know what I’m doing… shit might happen Saturday night.

Economically, many bookings are options purchases, not attendance forecasts — especially when the fee for no-show is trivial. The provider sees thirty bookings and plans for thirty bodies. The latent structure might be eighteen firm players, seven likelies, and five humans hedging. Binary systems destroy demand-quality information.

Restaurants have lived inside this fog for years. Reservation research treats no-shows as a first-class planning problem: studied sites show meaningful bands (on the order of low single digits to mid-teens in some samples; broader commentary often treats rates near twenty percent as unsurprising, with special occasions far higher).¹ Operators invent deposits and overbooking not because diners are uniquely immoral, but because a binary reservation is a soft commitment wearing a hard label.

An order book with elasticity

Imagine Sunday demand as a depth chart — not a list of booked names, but a stack of intent across times. Then add what binary software cannot see: each human’s elastic multi-time preference.

Scott: strong 10:00–10:30, fine at 11:00, outside at 12:00
Jane:  any time after 10:30, before 13:00
Peter: any Sunday morning if social/beginner-ish

That structure is an intent order book. The “prices” are times and conditions. The clearing problem is: which session, if synthesised, maximises compatible playable demand given real court constraints?

Current software does not solve that problem. It publishes a row and waits. A market maker’s job is to optimise the clearing time across flexible human intent — not to guess an 8 a.m. session and hope.

Market microstructure has vocabulary for what clubs accidentally built. A call auction batches orders and clears at a scheduled instant; a continuous market clears whenever sides cross.^2,3 The T−10-day inventory drop is a primitive call auction for recreational demand. Agents unlock continuous, demand-first clearing: standing graded intent pools until density crosses a threshold, then the venue is asked to stage.

The intent ladder

Replace BOOKED / NOT BOOKED with a type system the market can clear against:

INDICATIVE  →  LIKELY  →  FIRM  →  COMMITTED
(+ WEAKENING / CANCEL / OPTIONAL-conditional)

Indicative is standing preference without near-term world state. Likely means conditions look suitable. Firm means count me toward quorum. Committed means book/pay — and it should happen late enough to be true, not ten days early because the drop opened.

State transitions are evidence updates: sleep, weather, calendar, cancelled soccer, already-on-site. Side states keep the book honest. Never equate indicative with committed on dashboards executives trust.

Continuation intent

Wednesday: you play 5 p.m. Soccer cancels. At 6:30 you are already at the courts. You late-register for 7 p.m. Poker term. Eerily accurate.

Binary software knew only “booked 5 / not booked 7.” An agent can know you are on-site, the alternate plan died, and back-to-back midweek has precedent. That re-pricing is continuation intent — fresher, higher-quality demand than any form filled ten days prior. Season-long soccer pre-assembles liquidity with a forward contract; pickleball’s spot market re-auctions every session. Agents will not force everyone into seasons. They can make the spot market stop throwing away soft and continuation demand.

Inventory is an artefact

Deterministic transaction software grew up needing a database row before anyone could transact. So operators publish supply first and wait for demand to discover it. Inventory, in this sense, is an architectural scar — not a law of human desire.

Forty-seven zero-result Monday-night searches (illustrative scenario figure) are not proof that Monday demand is weak. They are evidence that the market never opened an object to transact against. Instrument negative search. Treat repeated holes as unformed supply.

Honesty cut: physical scarcity is real — finite courts, hosts, insurance. The polling ritual is the artefact. Synthesise the bookable session after intent densifies:

OLD:        inventory → discover demand
AI-NATIVE:  intent → assemble demand → synthesise inventory

That is the replace move applied to market design: do not merely automate discovery of the grid. Ask whether the grid should be the primary object at all.

Flash-mob markets (walkthrough)

A flash-mob market aggregates compatible latent intent until a quorum forms, then asks supply to stage.

1. Standing orders overnight — graded, multi-time, conditional intent in a travel radius. No session row yet.
2. World updates — rain eases, humans wake, agents firm mid-morning windows.
3. Quorum detection — firm-equivalent count clears policy for the venue format.
4. Contact supply — “Twenty-one probable players for 11:30. Can you staff and unlock?”
5. Synthesise session — the row is born as an output of clearing.
6. Firm humans — confirmations, not discovery chores; optional stimulation near threshold.
7. Play, dissolve — the scaffold was temporary. Liquidity was the product; the mob was the mechanism.

Human flash mobs fail because everyone must independently decide to show up. Agent flash mobs pre-load the cell: latent demand is warehoused as standing orders that sum past threshold before bodies move. Andrew Chen’s cold-start framing names the cousin idea: an atomic network — the smallest stable, self-sustaining network cell — then replicate adjacent cells.^4,5

The scarce resource moved

Industries change when the historical constraint that shaped them is removed — even if the customer need looks stable. Coordination, matching, reminders, and polling have been repriced. The old scarcity was the ability to organise thirty people. The new scarcity is thirty compatible people willing to show up at a compatible time.

SUPPLY POSTURE          DEMAND-ASSEMBLY POSTURE
create session    →     observe latent demand
please fill us    →     find density → form cohort
                        → acquire court → game

The court remains a resource. The game is formed demand. Soccer already solved liquidity with season contracts. Pickleball’s spot market must re-run the strategy analysis for agents — or keep charging humans attention tax to participate in dozens of tiny auctions a year.

When a customer complains about friction and an owner replies with screenshots proving the announcement existed, two altitudes collide. Screenshots answer UI. They do not answer “what business are you in?”

Same doctrine, different industry

Corporate learning cohorts re-run the same frame. Old scarcity: organise fifteen professionals, a facilitator, and a room via catalogue and email. New scarcity: fifteen compatible learners with aligned skill gaps, overlapping calendars, and willingness. Demand-assembly L&D pools graded learning intent, detects cohort density, synthesises the workshop when quorum clears, and contracts facilitation as supply response — rather than publishing quarterly calendars that run half-empty and calling it culture.

Restaurant reservations remain a useful stress test for the commitment bug (binary seats, no-shows, overbooking patches) even when the scarce-resource story differs.

Operator map

1. Implement the intent ladder (even manually) for your top session type.
2. Instrument zero-result and near-miss demand by window.
3. Write a quorum policy: synthesise vs stimulate vs wait.
4. Measure intent-to-game latency from density threshold to staged session.
5. Quarterly: “We used to be scarce at ___ / We are now scarce at ___.”
6. Answer ontology before screenshots.

Binary bookings were a reasonable interface for a world where software could not hold graded human intent and markets could not clear without pre-published rows. That world is ending in pieces.

You can spend the transition polishing the drop. Or you can build the book — elastic, graded, honest about scarcity — and assemble demand until the session is worth synthesising.

Aggregate first. Synthesise second. Stop forcing uncertain humans to lie early.