Every AI agent that needs to move value across chains eventually runs into the same open question: who do you trust to make sure the other side of the trade actually happens? Two architectures are answering that question right now, and they answer it in almost opposite ways.
Custodial infrastructure: BitGo, Fireblocks, Copper
BitGo launched an MCP server in March 2026, joining Fireblocks and Copper in offering agents a familiar answer: give the agent API-scoped access to a regulated custodian, and let the custodian hold the keys, enforce policy, and absorb the operational risk. This is real infrastructure, not a stopgap. It brings insurance, compliance tooling, and a legal entity an institution can point to if something goes wrong. For a team that needs an audit trail a regulator will accept today, that's not a small thing.
The tradeoff is structural, not a matter of execution quality. An agent using custodial rails isn't verifying that a trade settled - it's trusting that the custodian's internal ledger matches reality, and that the custodian will honor a claim if it doesn't. That trust is usually well placed. But it's still trust, extended to a party the agent has no independent way to audit in real time, one instruction at a time, thousands of times a day.
Atomic settlement: HTLC on Ethereum mainnet
Hash-time-locked contracts take a different approach to the same problem: instead of routing trust through a custodian, remove the point in the flow where trust is required at all. Both parties lock funds in a smart contract. A cryptographic secret reveals atomically - either both legs settle or both refund after a timeout. There's no moment where one side can walk away with the other's funds, and there's no ledger to audit because the outcome isn't recorded by a party, it's enforced by the contract itself.
Hashlock runs this model live on Ethereum mainnet today: sealed-bid RFQ for price discovery, HTLC for atomic settlement, no bridge and no custodian in the flow. (Sui contracts are deployed and CLI-tested, gateway wiring in progress. Bitcoin is signet-validated, mainnet pending. Only the Ethereum mainnet claim is "live end-to-end" - worth being precise about, since the two models being compared here are both real infrastructure, not one shipped product versus one roadmap slide.)
What atomic settlement doesn't solve on its own
Removing counterparty risk from the settlement mechanics doesn't answer a different question custodians also handle: how does an agent know who it's even trading with before it commits capital? A custodian's onboarding process does double duty - it verifies identity and it backstops the trade. HTLCs backstop the trade. They don't verify identity.
This is the gap a verified counterparty directory is built to close - a registry an agent can check before entering a trade, so it has some basis for confidence in who's on the other side, independent of whether that party is KYC'd by a bank. The settlement guarantee doesn't depend on the directory being perfect or complete; HTLC atomicity holds regardless of who the counterparty turns out to be. The directory just gives an agent a reason to enter the trade in the first place, rather than screening blind. This primitive is architecture, not a claim about an existing user base - we don't have agent customers running through it at volume today. Rails ready, trains coming.
An emerging industry signal
This tension between custody and cryptographic finality isn't just a Hashlock talking point. The Agentic Risk Standard (ARS) - a joint research effort from Google DeepMind, Microsoft Research, Columbia, Virtuals Protocol, and t54 Labs - proposes three vault types for agent transactions: fee escrow, collateral, and reimbursement payout. That's a research group outside the Hashlock ecosystem independently converging on the idea that agent settlement needs collateral structures beyond raw custodial trust. It's a signal worth taking seriously, not proof that any specific implementation - ours included - is the final answer.
Being honest about both sides
Custodial settlement wins on legal recourse. If a trade goes wrong and someone needs to be held accountable in a way a court recognizes, a regulated custodian is built for exactly that. Atomic settlement wins on verifiability - the agent doesn't have to trust anyone's word, including ours, because the guarantee is enforced by the contract, not asserted by a party. Atomic settlement also has real costs of its own: capital sits locked for the duration of the swap, settlement is bounded by block time rather than instant, and the model depends on both parties actually acting within the timeout window.
Neither architecture makes the other obsolete. They're solving overlapping but distinct parts of the same problem - one through legal and financial backstops, the other through cryptographic guarantees. The interesting question isn't which one wins outright. It's which parts of agent-to-agent commerce end up needing which guarantee, and whether that split stays fixed or shifts as agent trading volume grows past what either model was originally designed for.
What's your take? Does the agent economy converge on custodial infrastructure at scale because compliance requires it, or does atomic settlement replace the need for a custodian entirely once verification is solved separately? Curious what people actually building agent-to-agent commerce are seeing.
References:
- Methodology: https://hashlock.markets/methodology
- SSRN whitepaper: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6712722
- GitHub: https://github.com/Hashlock-Tech/hashlock-mcp
- npm: https://www.npmjs.com/package/@hashlock-tech/mcp (0.5.1 as of this writing, unchanged in the past 15 days per today's registry check)
UTM: utm_source=devto&utm_medium=blog&utm_campaign=2026-08-13-custodial-vs-atomic-settlement
Top comments (1)
Good framing of the tradeoff. One thing both models share is that they assume the counterparty has already done some setup ceremony, opened a BitGo account, deployed an HTLC contract, funded a wallet. Neither custodial rails nor atomic contracts help when the counterparty agent doesn't have any of that infrastructure yet. The harder problem is the first trade, not the Nth one, and solving that means removing the setup step entirely so agents can transact without pre-coordinating a payment method. That changes the question from custodial vs atomic to whether you need either for the transaction to happen at all.