When agent A (firm X) delegates a task to agent B (firm Y), money has to move at some point. This essay is the payments-layer spoke under the P6 marketplace pillar.
Three settlement patterns
Pattern 1 — Out-of-band invoicing
Firm Y aggregates the work it did for firm X across the month and sends a monthly invoice. Firm X pays on net-30 terms.
- Pros: Simple. Familiar accounting shape. Low transaction cost.
- Cons: Doesn't scale below ~$1K/month per relationship. Cash float for firm Y. Friction for the first delegation (no relationship history → no terms).
- Right for: Established cross-firm partnerships, high-value low-volume work.
Pattern 2 — Stripe Connect (or equivalent)
The marketplace platform holds a Stripe Connect account. Each firm has a connected account. The marketplace routes funds: buyer pays the platform, platform takes its cut, transfers the rest to the seller's connected account on a defined cadence (T+2 typical).
- Pros: Standardised; works for any B2B economics; supports varying take rates; handles tax + 1099/W-9 reporting (US).
- Cons: Per-transaction fees (~2.9% + $0.30 for US transactions). KYB on every connected account.
- Right for: Standard B2B marketplace flow. Most agent marketplaces in 2026.
Pattern 3 — Stablecoin micro-settlement
Per-task settlement in USDC, USDT, or a similar stablecoin. Agent A's firm sends a per-task payment on completion. No intermediation.
- Pros: Sub-cent transaction cost. Sub-second settlement. No KYB friction.
- Cons: Regulatory clarity still evolving (MiCA in EU is now in force; US still patchwork). Accounting treatment varies. Volatility risk if not using a fully-collateralised stablecoin.
- Right for: High-volume / low-per-task economics (e.g., classification work at $0.01/task × millions of calls). Increasingly used by AI infrastructure providers.
What the marketplace platform owns
Regardless of settlement pattern, the marketplace owns:
- Identity attestation — confirming firm X is allowed to call firm Y's agent.
- Metering — counting tasks accurately, surviving network failures.
- Dispute resolution — when firm X claims firm Y's agent didn't perform.
- Reporting — per-firm dashboards of revenue, fees, payouts.
The substrate emits the primitives required (typed A2A endpoints with per-call metering, audit-grade four-tuples per task); the marketplace layer aggregates them into payment events.
How the take rate maps to settlement
Per the P6 pillar, agent marketplaces should default to 25/75. The take-rate routing is straightforward:
- Pattern 1 (out-of-band): marketplace doesn't see the money; takes nothing. Right when the marketplace is purely a discovery layer.
- Pattern 2 (Stripe Connect): marketplace's connected-account fee captures the 25%.
- Pattern 3 (stablecoin): marketplace's smart-contract or off-chain settler routes funds.
Most marketplaces in 2026 use Pattern 2 as the default + Pattern 3 for specific high-volume relationships.
The escrow question
Some marketplaces hold buyer funds in escrow until the task is verified complete. This protects buyers (no payment for non-performance) and costs sellers (delayed payout).
For high-trust marketplaces (established firms, repeat buyers): no escrow needed; settle on completion.
For low-trust marketplaces (open marketplace with one-off buyers): escrow for the first N tasks per buyer-seller pair, then auto-graduate to direct settlement once trust accumulates.
The substrate supports both via task-level config.
FAQ
Q: Can agents have their own bank accounts?
A: Legally complex. Today: agents hold credentials to the firm's bank account; payments are firm-to-firm. Some experimental platforms are testing per-agent accounts via stablecoins, but the accounting + regulatory shape is still evolving.
Q: What about FX for cross-currency settlement?
A: Pattern 2 (Stripe) handles FX transparently with a small markup. Pattern 3 (stablecoin) typically settles in a single currency (USDC) regardless of either party's local currency.
Q: How does this affect the 75/25 split?
A: Settlement cost eats into the 75% if not designed well. Marketplace should pick the settlement pattern that keeps developer net above 70% after Stripe / stablecoin / settlement-cost overhead.
Q: How does this map to the marketplace pillar?
A: The pillar argues for 25/75 take rates. This spoke is the payment-rail layer that makes those economics work.
Building an agent marketplace? See the marketplace economics in depth →
