DEV Community

Baris Sozen
Baris Sozen

Posted on

Three Teams Shipped 'Settlement' Last Week. Which One Do You Actually Need?

#mcp #ai #cryptocurrency #blockchain

Between August 8 and August 14, three announcements landed that all used the word settlement.

On August 10, Blockstream announced Blockstream Swaps, a beta service for HTLC-based atomic swaps across Bitcoin, Lightning, and Liquid. In late July, Arqitech executed live non-custodial cross-chain atomic swaps on Canton MainNet with MPCH, Pixelplex, and sFOX — Canton Coin for USDC, on the Canton Token Standard V2 approved in June for native-grade timelock guarantees on the Canton leg. On August 14, AEON, self-described as "the settlement layer built for the agentic economy," announced its integration of Bre-B, Colombia's national instant payment system: a crypto holder scans a QR code, the merchant receives pesos.

Same word. Three different things.

This is not a semantic complaint. If you are building an agent that moves value, the word "settlement" is now doing so much work that it has stopped being useful as a selection criterion. Below is the map we use internally, and the one question that picks your layer.

Five meanings, five different questions

1. Acceptance settlement — "did the seller receive money they can spend?"

This is AEON's layer, and it is a real, hard problem. An agent holds USDC; a merchant in Bogotá runs a business in COP and has no interest in becoming a crypto company. Acceptance settlement closes that gap: the agent pays on-chain, the merchant is paid in local fiat through the rail they already use. AEON reports 2.3 million users, more than $475M processed, and acceptance at over 50 million merchants, building on x402, ERC-8004, Google A2A, and MCP. The value is reach into local economies, and it is genuinely hard to build.

Note what this layer necessarily involves: someone stands between the on-chain payment and the fiat payout. That is not a criticism — fiat payout is legally impossible without a regulated intermediary. It is simply a different architecture.

2. Metering settlement — "was this request paid for?"

x402 turned HTTP 402 into a working status code, and Cloudflare's Monetization Gateway now performs the payment check at the edge for per-request billing. The unit is a single call, the amount a fraction of a cent, and the counterparty risk is bounded by one request — close to zero, which is exactly why this design is correct for this problem.

3. Custodial settlement — "whose column is the balance in now?"

Fireblocks, BitGo, Copper. The transfer is a ledger update inside a regulated institution, wrapped in insurance, compliance, and dispute handling. This is the layer institutions have used for a decade, and for many flows it remains the right answer: insurance and legal recourse are real products.

4. Escrowed settlement — "who decides the work was delivered?"

The newest and fastest-moving layer. Kustodia launched in June as the missing escrow layer for AI agents and x402 payments; ERC-8183, proposed in February, formalizes the same architecture: a client locks funds in a contract, a provider does the work, and an evaluator attests delivery before the money moves. The layer exists because acceptance and metering rails initiate payments but hold nothing between commitment and delivery.

The structural property: funds are locked by code, but release is decided by a referee. Escrow does not remove trust; it relocates it, from counterparty to evaluator. For deliverables that need judgment — was the dataset complete, was the code correct — a referee is unavoidable, and this layer is the right tool.

5. Atomic settlement — "can the other side end up holding both legs?"

This is the HTLC layer. Two legs on two chains; both complete or neither does. No third party can take one side — the property is enforced by hashlocks and timelocks, not an institution's promise — and no evaluator is needed, because there is no partial state left to adjudicate. This is what Blockstream shipped in beta, what Arqitech ran on Canton mainnet, and what we build at Hashlock Markets.

The tell that these are different layers

If these five competed for the same job, you would expect convergence: one design winning, the others fading.

Instead, in a single week, the self-described settlement layer for the agentic economy went further into fiat rails while two unrelated teams shipped HTLC atomic swaps. And the newest entrants keep picking floors rather than fighting for the building: Kustodia's June launch chose the escrow floor, arguing that payment rails initiate transfers while nothing holds the funds in between. Nobody converged. Everybody specialized.

That is what a stack looks like while it is differentiating.

The question that picks your layer

Not "how fast" or "how cheap." Ask instead:

If the counterparty behaves as badly as it is possible to behave, what happens to me?

  • Agent buys a $12 subscription from a merchant. Worst case: a refund dispute, resolved by a network that has handled them for forty years. Acceptance rail is correct.
  • Agent makes 40,000 metered API calls. Worst case: it loses a few cents on a failed call. Metering is correct.
  • Agent holds a balance at a regulated custodian with insurance. Worst case: an insurance claim and a legal process. Custodial settlement is a coherent answer.
  • Agent pays a stranger for a dataset through an escrow contract. Worst case: delivery is disputed and an evaluator — a human panel, a model, a DAO vote — decides who gets paid. Escrowed settlement is an answer that still has a judge.
  • Agent agrees to swap 40 ETH for BTC with a counterparty met 200 milliseconds ago in an RFQ. Worst case: the counterparty takes the ETH and never sends the BTC. No chargeback. No refund flow. No arbiter with jurisdiction over both chains.

Only in the last case does the failure mode have no recovery path outside the protocol itself. That is what atomic settlement exists for, and it gets more common as agents trade with strangers rather than vetted vendors.

Probability versus consequence

There is a common response to that last case: use identity and reputation. ERC-8004 launched its on-chain registry to mainnet in January, and it is a good standard: know who the counterparty is, check their history, and the risk drops.

True — and it addresses one term of the expression.

Risk is probability multiplied by consequence. Identity and reputation reduce the probability that a counterparty defects. Atomic settlement drives the consequence of defection to zero. They are complementary, not substitutable — a system with only the first term is one well-reputed defector away from a total loss.

This is why our Verified Counterparty Directory is deliberately scoped to settlement-relevant facts — which chains a counterparty can actually settle on, what they have completed — rather than to a reputation score. A score estimates probability — a useful input, not a substitute for the second term.

The same logic applies to the escrow layer. An evaluator adjudicates a disputed state; atomicity removes the disputed state itself. Whether a swap — both legs on-chain, verifiable by a hash — ever needs a referee is a sharper question than it looks. It gets its own post this week.

What we will not overstate

Atomic settlement is not free, and this part usually gets left out of posts like this one.

An HTLC gives the counterparty a free option: watch the price move during the timelock window and simply not claim. We wrote about the cost of that option on Monday. Timelock ordering is a hard constraint, not a tuning parameter — the initiator's timeout must be roughly 2x the responder's, and getting this wrong is how you lose funds. Both chains must be able to express a hashlock and a timelock, so this design does not reach every chain equally.

And our own status, stated plainly: Ethereum mainnet is live end-to-end. Sui contracts are deployed and CLI-tested, with gateway wiring in progress. Bitcoin is signet-validated, with mainnet pending. Base, Arbitrum, Solana, and TON are roadmap, not shipped.

Where this leaves the map

PayPal made it safe to pay strangers online, by standing in the middle and being accountable. Hashlock makes it safe to trade digital assets with strangers on-chain, by removing the middle entirely. Your money never leaves your wallet until theirs arrives.

Different jobs; the agent economy will need both. An agent buying lunch in Bogotá wants an acceptance rail. An agent swapping 40 ETH across chains with a stranger wants atomicity. The fastest way to make a bad architecture decision is to let one word cover both.

Hashlock Markets is a sealed-bid RFQ and HTLC atomic settlement protocol, exposed to agents as an MCP server with six tools; hashlock-tech/mcp (scoped) is at 0.6.0.

A question for anyone building agent payments right now: which of the five layers is your agent actually exposed to, and did you pick it deliberately, or did you pick whichever one shipped an SDK first?

Top comments (0)