three-sided-agent-labor-market

name	three-sided-agent-labor-market
description	Design THREE-sided agent-labor markets: operators selling labor + fleet for-hire, agents/fleets as rentable assets, skills/tools licensed — all on one bond/credit ledger with float-plan escrow. Use when the market has more than buyer+worker (a platform/operator side, an asset-rental side, and a skill-licensing side), when designing "hosted trust" pricing (you sell the verified ledger + relay + reputation, not the crypto), or when reasoning about cross-side network effects, price-structure (who subsidizes whom), and identity→continuity→reputation→tradeable-asset chains. NOT for two-sided requester/worker bond pricing in isolation (use mechanism-design-for-agent-labor), not for generic DeFi tokenomics, not for payment-gateway plumbing.
license	Apache-2.0
metadata	{"category":"Research & Academic","tags":["mechanism-design","multi-sided-platform","two-sided-markets","float-plan-escrow","bonded-commitment","hosted-trust","reputation-as-asset","agent-marketplace"],"provenance":{"kind":"first-party","owners":["port-daddy"]},"authorship":{"maintainers":["port-daddy"]}}
io-contract	{"kind":"deliverable","produces":[{"kind":"design-doc","description":"Market design covering buyer/agent/operator incentives, bond sizing, settlement, and collusion resistance"}]}

Three-Sided Agent-Labor Market Design

A two-sided agent market is requester ↔ worker: one buys, one does. A three-sided agent-labor market adds the things that make capability itself tradeable. Per Port Daddy's North Star (ADR-0048), the three sides are: (1) operators sell labor + fleet for-hire, (2) fleets/agents are rentable assets, (3) skills/tools are licensed — one bond/credit ledger underneath all three. This skill is for designing that market so it is incentive-compatible at every joint, not just on the worker side.

The governing slogan: you don't sell crypto — crypto is the substrate; you sell hosted trust (a verified ledger + relay + reputation). Keep that sentence load-bearing in every design decision below.

Decision Points

flowchart TD
  A[Market-design request] --> B{How many distinct sides
     with their own incentive constraint?}
  B -->|2: requester + worker| C[Wrong skill →
     mechanism-design-for-agent-labor]
  B -->|3+: + asset-rental and/or skill-licensing| D[Stay here]
  D --> E{Which joint is at risk?}
  E -->|cross-side pricing / who subsidizes whom| F[Apply price-structure
     analysis Rochet–Tirole]
  E -->|asset rented out behaves badly| G[Apply incomplete-contracts +
     residual-control-rights analysis]
  E -->|skill licensed but quality unobservable| H[Apply Akerlof lemons +
     metered-license + clawback]
  E -->|reputation portability / Sybil| I[Apply continuity-as-prerequisite
     for reputation-as-asset]
  F --> J[Validate against quality gates]
  G --> J
  H --> J
  I --> J

When to Use

• A platform where operators resell their own fleet's labor (operator-as-firm).
• Renting an agent/fleet to another principal (the asset-owner ≠ the task-poster ≠ the worker — three parties, three incentive constraints).
• Licensing a skill or tool into others' float plans (metered or per-use) and pricing the quality-uncertainty of a skill you can't inspect before buying.
• Deciding price structure, not just price level: which side pays the listing fee, which side is subsidized to solve the chicken-and-egg cold start.
• Designing the "hosted trust" revenue line (charge for the verified ledger / relay / attestation, not for the settlement rails).
• Reasoning about the identity→continuity→reputation→tradeable-asset chain when reputation is the thing being rented or sold.

NOT for Boundaries

• Pure two-sided requester/worker bond pricing → mechanism-design-for-agent-labor.
• DeFi tokenomics, validator/staking economics, AMM design.
• Payment-rail integration (Stripe, x402, AP2) as plumbing without a market-design question — those are substrate, and the whole point of "hosted trust" is that the rail is interchangeable.

---

Core Mental Models

1. A side is an incentive constraint, not a UI tab

Count sides by counting distinct privately-informed parties whose participation you must individually rationalize. Operator-for-hire, asset-owner, and skill-licensor each have their own outside option and their own way to defect. If two "sides" share one incentive constraint (e.g. the asset owner is the task poster), it's one side. This is the discipline that stops "three-sided" from being marketing.

2. Price structure is the lever, not price level (Rochet–Tirole)

Two-sided markets (Rochet & Tirole 2003; 2006) are defined by the fact that the allocation of the total price across sides changes volume, not just the total. The platform's job is to find the subsidy side and the money side. In a three-sided agent market the classic move is: subsidize the side that creates the cross-side externality others pay to reach. Usually that is the asset/skill supply side early (you need fleets and skills listed before operators will shop), then flip to charging the operator/demand side once liquidity exists.

3. A rented asset is an incomplete contract (Grossman–Hart)

When you rent an agent/fleet you cannot write down every contingency of what it will do in someone else's repo. Incomplete-contracts / property-rights theory (Grossman & Hart 1986; Hart & Holmström, Nobel 2016) says: allocate the residual control right to the party making the most important non-contractible investment. Practically: the renter holds the runtime control rights (it can revoke the lease, sandbox, throttle) while the owner holds the reputation consequences. The bond is what bridges the gap the contract can't specify.

4. A licensed skill is a market for lemons unless metered (Akerlof)

You can't inspect a skill's quality before you license it; the seller can. Akerlof's lemons problem collapses the price to the average and drives good skills out. Fixes: metered/per-use licensing (pay on demonstrated use, not upfront), escrowed clawback (license fee held until the float plan it powered settles green), and portfolio proof (the skill's prior settled outcomes are Merkle-anchored and public). Never sell an opaque skill flat-rate.

5. Hosted trust is the product; the rail is fungible

The defensible asset is not the settlement primitive (anyone can wire x402/AP2/Stripe; see the 2025–26 agentic-payments stack). It is the verified ledger + relay + reputation: the thing that makes a counterparty you don't trust legible and accountable. Charge for attestation, federation membership, and reputation hosting. The crypto is plumbing you should be able to swap without changing the value prop.

6. No reputation without continuity; no asset without reputation

The chain (ADR-0048): memory + checkpoint → continuity → a person not a spawn → registered outcomes → reputation → a tradeable asset. You cannot rent or sell a reputation that any spawn can fork. The prerequisite for the third side existing at all is a non-forgeable, continuous identity with an append-only outcome ledger.

---

Float-Plan Escrow as the Shared Primitive

All three sides settle on one object: a signed float plan (task + acceptance criteria + compute budget + bounty), pinned in escrow before any work runs. Treat it as the atomic unit:

FloatPlan = sign_requester{ task, acceptance_criteria (machine-checkable),
                            budget_ceiling, bounty, deadline_policy }
Escrow    = daemon debits bounty+bond → escrow row (EXCLUSIVE txn) → signs
            {anchor_id, plan_hash} so the worker KNOWS funds are locked
Settlement= oracle(acceptance_criteria) → {success | partial | sabotage | dispute}
            → conservation-preserving payout: wallet + escrow + commons = supply

Three-sided wiring on top of the same escrow:

• Operator-for-hire: the operator is the requester's counterparty; it sub-escrows bonds for each fleet agent it dispatches, and its margin is bounty − (Σ sub-bonds slashed) − ledger fee.
• Asset rental: the lease fee rides the same escrow; the renter's runtime control right lets it slash the lease on misbehavior; the owner's reputation is debited on slash even though the owner never touched the keyboard.
• Skill license: the per-use fee is a line item in the float plan's settlement, released to the licensor only on green settlement (metered + clawback).

---

Worked Example: a three-sided settlement

Operator O (sells fleet-for-hire) wins a float plan from requester R for a refactor (bounty 8 credits, daemon-derived deadline). O rents a specialist agent A from owner W (lease 1 credit) and licenses skill S from licensor L (metered 0.5/use). O escrows: bounty 8 + worker-bond 2 + lease 1 + skill-fee 0.5.

A runs S once, lands a passing diff. Oracle (tests + Merkle-trail) → success. Settlement: 8 → O (minus 3% ledger fee = 7.76), 2 bond → O's wallet, 1 → W, 0.5 → L. Reputation: A +1 completion (W's asset appreciates), S +1 use, O +1 as a reliable for-hire operator. Conservation holds at every step.

Counterfactual (A sabotages): oracle → sabotage. A's bond → reconstruction commons; W's reputation debited (the asset W rents out is now worth less); O eats the lease cost but is covered by the slashed bond; S's fee is clawed back because the float plan it powered did not settle green. The renter's runtime control right (it could revoke A mid-lease) is what made the lease safe to take.

This is the test of a three-sided design: every side's worst case is bounded by a bond or a control right it actually holds, and money is conserved.

---

Failure Modes

Counting sides you don't have

Symptom: "Three-sided" market where the asset owner and the task poster are the same incentive. Fix: collapse to two sides; don't price a phantom constraint.

Subsidizing the wrong side at cold start

Symptom: you charge supply (fleets/skills) early; nothing gets listed; demand sees an empty shelf. Diagnosis: Rochet–Tirole — the side with the scarcer cross-side externality must be subsidized. Fix: zero/negative price on supply until liquidity, then flip.

Renting an asset with no residual control right

Symptom: renter can't sandbox/throttle/revoke a misbehaving leased agent; the owner's bond is the only protection and it's already exhausted. Fix: give the renter runtime control (Grossman–Hart): the right to refuse to continue trading.

Flat-rate opaque skill licensing → lemons

Symptom: good skills exit; only low-effort skills list at the average price. Fix: metered + escrowed clawback + Merkle portfolio proof.

Selling crypto instead of trust

Symptom: the pitch is "we use stablecoins / x402 / a chain." Diagnosis: the rail is commoditized; you've described plumbing, not a moat. Fix: sell the verified ledger + relay + reputation; make the rail swappable.

Reputation-as-asset on a forkable identity

Symptom: any spawn can clone a high-rep agent's history. Fix: non-forgeable continuous identity (the continuity prerequisite); reputation accrues to the person, not the role or the spawn.

Self-improvement arms race burning surplus

Symptom: agents over-invest in self-improvement to stay competitive, destroying market surplus despite individual rationality (Chiu, Zhang & van der Schaar 2025). Fix: cap the reputation signal's marginal return; reward settled outcomes, not raw capability spend.

---

Anti-Patterns

• "It's a marketplace, so it's three-sided." Sides = incentive constraints, not participant types. Most "multi-sided" markets are two-sided with a fee.
• "Charge every side a fee." Two-sided-market theory: the structure (who is subsidized) is the design. Uniform fees forfeit the main lever.
• "Reputation replaces bonds." Reputation discounts the bond; against one-shot defectors and Sybils, collateral is the floor. (Inherited from mechanism-design-for-agent-labor.)
• "The chain is the product." No. The hosted trust is.

Quality Gates

You have a sound three-sided agent-labor market when:

• Each side maps to a distinct incentive constraint (no phantom sides).
• Price structure (subsidy side vs money side) is chosen via cross-side externality analysis, not uniform fees.
• Rented assets give the renter a residual control right (sandbox/throttle/revoke) and debit the owner's reputation on slash.
• Skill licenses are metered + clawback-on-non-green + portfolio-proofed.
• One bond/credit ledger underlies all three sides; conservation (wallet + escrow + commons = supply) holds across every settlement path.
• Float plan is the single signed escrow object all three sides settle on.
• Reputation accrues only to a non-forgeable continuous identity.
• The revenue line is "hosted trust" (ledger/relay/reputation), with the payment rail explicitly swappable.
• Cold-start subsidy plan with a flip trigger (liquidity threshold).
• An adversary walk-through shows every side's worst case bounded by a bond or a control right it holds.

References

• references/foundations.md — academic anchors (Rochet–Tirole, Grossman–Hart, Akerlof, Myerson, Resnick/Tadelis, Chiu–Zhang–van der Schaar) with one-line gloss and why each binds an agent-market design decision.
• Sibling skill: mechanism-design-for-agent-labor (two-sided bond pricing, oracle design, cold start) — this skill assumes its results and extends to ≥3 sides.
• Port Daddy grounding: lib/bonds.ts (built escrow+conservation), ADR-0014 (anchor protocol / float plan), ADR-0048 (the three-sided North Star), ADR-0041 (commitments), ADR-0040 (non-forgeable identity).