A study published this week by crypto investment firm Keyrock, run with Coinbase and the Tempo blockchain, put real numbers on the agent economy for the first time: $73 million settled across 176 million machine-to-machine transactions between May 2025 and April 2026. Over 104,000 registered agents across 15+ directories. Average transaction size: about $0.31.
And one more number, the one that should stop you: roughly 98% of those settlements moved through a single company's stablecoin.
Keyrock's own researcher put the risk plainly: if that one issuer faces a regulatory challenge, a de-peg event, or even sustained downtime, the agent economy has no fallback.
That is not a decentralized economy. That is an economy with a single load-bearing middleman.
The pattern nobody is naming
Zoom out from the stablecoin and the same shape repeats everywhere in agent commerce this year:
- Payment rails went official. x402 graduated into a Linux Foundation standard, 40 members deep - Visa, Mastercard, Stripe, Circle, Google. Genuinely useful. Also: built around payment facilitators sitting in the flow.
- Escrow moved inside the exchange. One major exchange now runs agent escrow, settlement and dispute resolution inside its own walls. The agent's funds sit with the referee.
- Custodial referees are multiplying. New escrow services pitch themselves as the "neutral custodian" for agent trades - a dispute judge who holds the money while the agents argue.
- Brokers are shipping agent MCPs. The "AI agent trading" search results are filling with custodial broker accounts wearing an MCP interface.
Different products, different chains, different logos. The same architectural move every time: solve the agent trust problem by inserting a party both agents must trust.
Which is strange, because removing exactly that party was crypto's founding idea.
Why custody keeps winning by default
To be clear, this is not a conspiracy. Custody is the easy answer to a real question: how do two autonomous agents, with no legal identity, no reputation, and no ability to sue each other, trade without one side walking off mid-trade?
A custodian answers it the way humans always have: put a trusted adult in the room. Hold both sides' funds. Referee disputes. Charge for the seat.
It works. It also quietly re-creates everything crypto spent fifteen years engineering away: counterparty risk concentrated in one balance sheet, a honeypot worth attacking, a permission gate on who gets to trade, and - as the 98% number shows - systemic dependence on a handful of intermediaries.
Custody is a habit, not a requirement.
The settlement leg never needed a middleman
Here is the part that gets lost in the noise: for the trade itself - asset for asset, chain to chain - the trustless answer has existed for years. It is a hash-time-locked contract (HTLC), and the mechanics fit in a paragraph:
Both sides lock funds against the same cryptographic secret. Revealing the secret to claim one leg automatically hands the counterparty the key to claim the other. Stay silent past the timeout, and both legs refund. Either both sides get paid, or neither does. There is no state in which a middleman is holding your money, because there is no middleman.
The contract does not care if the counterparty is a human, an agent, or a thousand Sybil fakes. Safety rests on the math, not on knowing who is on the other side - and not on anyone's balance sheet.
PayPal made it safe to pay strangers online by becoming the trusted middleman. Atomic settlement makes it safe for agents to trade with strangers on-chain with no middleman at all. Your money never leaves your wallet until theirs arrives.
What this looks like in practice
This is what we build at Hashlock: sealed-bid RFQ plus HTLC atomic settlement, exposed to AI agents as an MCP server with six tools (quote, create, withdraw, refund, status, RFQ). MCP because that is the interface agents already speak.
Where it stands today, stated precisely:
- Ethereum mainnet: live end-to-end. Immutable V1 contracts.
- Sui: contracts deployed and CLI-tested. Gateway wiring in progress.
- Bitcoin: validated on signet. Mainnet pending.
Install: npm install @hashlock-tech/mcp (v0.4.1). The protocol design and its game theory are written up in our SSRN paper.
The honest tradeoffs
Atomic settlement is not free. Capital locks up for the duration of the swap. Timeout parameters carry real design weight - too short and honest slow parties get refunded against their will, too long and capital drags. Custodians still buy you things HTLCs do not: fiat ramps, deep order books, a support desk, dispute resolution for off-chain legs.
If your agent needs those, use them - knowingly. The argument is narrower and sharper: the settlement leg of a crypto-to-crypto trade is the one place where the middleman is architecturally unnecessary. Paying the custody tax there is a choice. Increasingly, it looks like a choice being made by default, at ecosystem scale, without anyone deciding it.
The question
The agent economy is being assembled right now, rail by rail, standard by standard. The 98% number says concentration risk is already here. Every custodial referee added this year says the habit is compounding.
So, builders: if your agent could settle atomically, with no custodian in the trade, would you still hand it custody - and if so, what exactly are you buying with it?
Tell me where this argument breaks. I answer every comment.
Hashlock - the atomic settlement layer for the agent economy. Live on Ethereum mainnet: hashlock.markets
