A piece in CoinDesk this week made an argument I have been thinking about for two years: AI agents could finally solve crypto's user-experience problem. Not by wallpapering over it. By being the user.
It is hard to disagree. A self-driving treasury that never types a seed phrase, never copies a contract address, never accidentally bridges to the wrong chain — that is the crypto UX everyone gave up waiting for. The 1B+ Haun Ventures fund announced three weeks ago, the 25M Portal to Bitcoin raise, the 8.3M for Nava — capital is moving on this thesis, fast.
The thesis is right. The stack underneath it is not finished. There is one layer the agent-payment standards have not built, and without it the most interesting class of agent action — placing and settling a real trade — remains stuck behind the same trust assumptions that defined OTC desks in 2018.
This is the map of where the autonomous economy is, and where the missing piece sits.
What has actually shipped for agents in the last six weeks
By rough count, five things now exist that did not exist on April 1.
ERC-8004 ratified earlier this year is the agent-identity standard on Ethereum mainnet. It gives an agent a stable, verifiable, on-chain "who I am."
ERC-8183 is the agent-commerce standard, co-developed by the Ethereum Foundation and Virtuals Protocol. It defines how an agent expresses commercial intent — a thing it wants to buy, a delivery target, a payment commitment.
x402 is Coinbase's HTTP-payment protocol. Last published numbers: ~167M cumulative transactions, 3.3x growth in roughly two weeks, 69k+ agent endpoints, the x402 Foundation now sits inside the Linux Foundation.
OKX Agent Payments Protocol launched late April — agent-to-merchant payment rails, with an "escrow coming soon" badge that has been there since the announcement.
Pay.sh went live early May with Solana Foundation and Google Cloud — x402-compatible, integrated with Gemini, BigQuery, and Vertex AI.
That is a lot of stack to ship in six weeks. It is also, in a meaningful sense, all the same shape: an agent paying a service for a thing it consumed, in a single direction, in a single moment.
The thing that still doesn't exist
Settling a trade is not paying for an API call.
A payment is one-sided. You owe me a fixed amount; I deliver a service; you transfer the amount; we are done. The protocol can be T+0, the lock can be a few seconds, and the only failure mode is "did the payment land or not."
A trade has two sides, and usually two chains. I owe you USDC on Ethereum. You owe me BTC on Bitcoin. The atomic guarantee everyone wants — both legs settle, or neither does — is mathematically impossible without something stronger than a payment rail. You need either a custodian who holds both sides of the float (the centralized OTC desk model, with all of its 2018 baggage) or a cross-chain primitive that lets two parties commit funds on two chains and reveal-or-refund them under a shared timelock.
That primitive is the Hashed Timelock Contract, and it is about thirty years old. The interesting work has been making it usable for two-sided OTC at agent latencies, on chains where the contract is the escrow.
That is the layer none of the five payment standards builds. It is the layer the next class of autonomous-economy actions — sell BTC, buy ETH, hedge a forward, rebalance a treasury across three chains — needs underneath it.
The five primitives we shipped to fill it
Hashlock Markets has been shipping into this gap for six weeks. As of this Sunday the surface is five primitives, all callable via the same MCP server from any agent runtime.
BTC Collateral Vaults — your Bitcoin stays on Bitcoin. A cryptographic balance proof unlocks the Sui-side leg of a trade via Hashi's Sui-native bridging. There is no wrapper, no synthetic, no custodial IOU. Native BTC, native settlement. Yesterday's piece on Dev.to walked through the mechanics.
Forward OTC Settlement — a sealed-bid RFQ produces a price, an HTLC with a 24-hour or 48-hour timelock holds both sides' funds, and the clock decides the outcome. The five payment standards are all T+0; this is the first on-chain primitive that lets an agent commit today to a trade that delivers tomorrow without trusting a counterparty to honor the forward.
Verified Counterparty Directory — agents and human operators can register with tiered KYC, accumulating an on-chain reputation tied to their settlement history. A counterparty who has cleared 200 trades over six months at 98% on-time is structurally different from a fresh address. The directory makes that legible to a sealed-bid auction.
Multi-leg Trade Atomicity — sell BTC on Bitcoin, buy ETH on Ethereum, buy SUI on Sui — three legs across three chains, and either all of them clear or none of them does. The HTLC pattern generalizes beyond two-sided pairs. The intent travels as a single object; the contracts enforce all-or-nothing across the legs.
Execution Rewards and Tiered KYC — the protocol fee is 7 bps deployed, with up to a 50% rebate for participants who settle on time and meet the directory's reputation tier. Effective fee floor: 3.5-7 bps. For comparison, conventional CEX OTC desks usually quote 8-10 bps before custodial spread.
These five primitives are not aspirational. The contracts are deployed and immutable on V1; the audit RFP went out last week; the chain support is ETH, BTC, and Sui live, with Base, Arbitrum, Solana, and TON on the roadmap.
How they compose
Here is what the full settlement stack looks like to an agent on a real trade:
- The agent expresses intent — "sell 0.5 BTC, buy 7 ETH, settle within 24 hours."
- A sealed-bid RFQ goes to the Verified Counterparty Directory; quotes return privately.
- The agent picks a quote; the protocol opens HTLCs on Bitcoin and Ethereum with a T+24h timelock.
- Both sides escrow into the contracts. The float is in the contracts, not on a desk balance sheet.
- At settlement, the preimage is revealed; both legs clear atomically. If either side defaults, the timelock returns the funds.
- The on-time settlement triggers a fee rebate; the counterparty's reputation score in the directory updates.
Five primitives, one MCP surface, six tools. The same six tools any agent runtime can call: create_rfq, respond_rfq, create_htlc, withdraw_htlc, refund_htlc, get_htlc.
What this is not
It is not production-grade volume yet. The reference deployments are pre-audit, the V1 contracts are deliberately immutable so that no one — including us — can hot-patch them, and meaningful liquidity will follow the audit firm sign-off and the V2 deploy. The fee numbers above are the deployed schedule, not aspirational.
It is also not in competition with x402, Pay.sh, OKX APP, or ERC-8183. Those rails settle payments. Hashlock settles trades. The five payment standards and the trustless-settlement layer are complements; an agent treasury that runs on autopilot in 2026 needs both.
The next six weeks
Two things determine whether this stack matures into the rail an agent treasury can actually live on.
The first is the audit. V2 mainnet HTLC contracts are gated on external audit firm sign-off; the RFP is live and the timeline is weeks, not months.
The second is the chain expansion. Sui mainnet went live nine days ago; Base, Arbitrum, Solana, and TON are queued. Each new chain widens the multi-leg primitive's reach — a treasury that wants to rebalance across five chains is one chain shy today.
Past that, the thesis from the CoinDesk piece holds. Agents are the user crypto needed. Trustless settlement is the layer the agents need.
Which layer is still missing from your stack?
Stack: ETH, BTC, Sui live | Base, Arbitrum, Solana, TON on roadmap.
Spec, contracts, and MCP server: https://hashlock.markets/docs
Source on GitHub: https://github.com/Hashlock-Tech/hashlock-mcp
Full protocol spec on SSRN: papers.ssrn.com/sol3/papers.cfm?abstract_id=6712722
Top comments (0)