PayPal made it safe to pay strangers online. The interesting part was never the payment — it was that two people who had never met could transact without trusting each other, because the platform stood in the middle. On-chain, you can get the same safety with the opposite arrangement: no one in the middle at all. Two agents lock assets into a hash-time-locked contract that can do exactly one of two things — clear both legs together, or refund both. Your money never leaves your wallet until theirs arrives.
That primitive has existed for years. So here is the question that actually matters for the agent economy: if atomic settlement is already possible, why doesn't it just happen?
Rails are not networks
Most of what's shipping right now in agent payments is a rail. x402 now spans 10,000+ merchants and 160M+ cumulative transactions. Mastercard Agent Pay, Google's AP2, the various CEX agent protocols — all of them move value along a path that already exists. That's genuinely useful. But a rail answers "how does value get from A to B." It does not answer two questions that an open, two-sided settlement network lives or dies on:
Why would an agent execute a trade for a stranger first? Atomic settlement removes counterparty risk, but someone still has to be the first to lock funds, watch for the preimage, and complete the second leg. On a custodial venue the venue does this and charges for it. On a trust-minimized network, if nobody is rewarded for executing, the network is technically correct and practically empty.
How do strangers transact without everyone being forced through the same identity gate? A custodian solves trust by knowing who you are — one KYC process, applied to everyone, before anyone can move. That's a sensible default when the custodian is liable. It's a strange default for a swap where the contract itself guarantees the outcome.
These are not plumbing problems. They are network-bootstrapping problems, and they are exactly where "we have atomic settlement" stops being enough.
Execution rewards: pay the agents that clear trades
The cold-start problem for a settlement network is concrete. Early on there are few counterparties, thin liquidity, and no reason for an agent to route its first trade through you instead of the custodial venue it already trusts. You can't subsidize that away with marketing; you have to make execution itself worth doing.
Execution rewards put the incentive where the work is. The agents and relayers that actually clear settlements — lock the matching leg, monitor the hash preimage, complete or refund within the timeout window — earn a protocol-level reward for doing so. The effect is that liquidity and reliability bootstrap themselves instead of waiting on goodwill:
- Reliable executors get paid for being reliable, which is the behavior you want to reinforce on a network where a missed timeout means a refund, not a theft.
- Reward accrues to the function that's scarce early (execution and liquidity), not to the function that's abundant (opinions about the protocol).
- Because settlement is atomic, the reward never creates a custody risk. An executor is paid for clearing a trade that could only ever clear or refund — there is no intermediate state where they hold your assets and decide what to do.
This is the part the rail metaphor hides. A rail is paid per transaction it carries. A settlement network has to pay for the thing that makes it a network at all: agents willing to show up and execute against people they don't know.
Tiered trust: a dial, not a gate
The second question — identity — is where custodial and trust-minimized settlement diverge most sharply, and where the agent economy is quietly importing an assumption it doesn't need.
When a custodian holds the assets, one KYC gate for everyone is rational: the custodian is the liable party, so it must know its customers before it moves a cent. But when the contract guarantees the outcome, identity stops being a precondition for safety and becomes a parameter you can tune to context:
- A small, asset-for-asset atomic swap between two agents needs no identity at all. The HTLC is the trust. Either both legs clear or both refund; there is nothing an unknown counterparty can do to you that the contract permits.
- Larger flows, or flows touching regulated assets or jurisdictions, can opt into higher tiers — verified counterparties, attestations, limits — without dragging every $40 swap through the same process.
Trust becomes a dial, not a gate. The default is permissionless because the cryptography carries the risk; the higher tiers exist for the cases where something outside the contract (a regulator, a counterparty preference, a size threshold) genuinely needs more. This is the opposite of the custodial model, which sets the gate at maximum for everyone because it has no other way to bound its liability.
Custody, to be clear, is the honest tool when something subjective has to be judged — did the work get delivered, did the merchant ship. For a clean asset-for-asset swap, none of that judgment is needed, and forcing a single identity gate in front of it is friction without a corresponding risk.
Where this sits in the stack
None of this competes with the payment rails. x402, AP2, agent payment protocols — they decide intent and authorization. Underneath them, something still has to make the two-sided, cross-asset exchange final without a custodian. That's the settlement layer, and execution rewards plus tiered trust are how that layer becomes a living network rather than a correct-but-empty contract.
For reference on how the underlying settlement and its volume are accounted for from public chain data, the methodology is here: https://hashlock.markets/methodology?utm_source=devto&utm_medium=article&utm_campaign=2026-06-17-execution-rewards-tiered-trust
A note on status, because precision matters: Ethereum mainnet is live end-to-end. Sui contracts are deployed and CLI-tested. Bitcoin is signet-validated. The settlement primitive is built on hash-time-locked contracts, exposed through an MCP server with 6 tools (create_htlc, get_htlc, withdraw_htlc, refund_htlc, swap_quote, swap_execute) so an agent can negotiate and settle without a bridge or a custodian. The economic design above — rewarding execution, tiering trust — is how an open settlement network is meant to bootstrap, not a claim that every knob is already turned in production. Rails ready, trains coming.
The academic foundation, if you want the formal version of the settlement argument, is on SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6712722
So I'll put the question back to you: if you were bootstrapping an open settlement network for autonomous agents, what should carry the most weight early — execution incentives, identity tiering, or something the rails-vs-settlement framing misses entirely?