Agents can't keep secrets
BlackRock put out a paper last week called The Machine Native Economy. Agents buying their own compute, their own data, their own API calls, paying in stablecoins over crypto rails, with compute itself becoming collateral you can borrow against. IDC, the research firm, says a billion agents will be live by 2029, doing 217 billion actions a day. Gartner says machines will steer $30t of purchases by 2030. BCG says $16t of assets will be tokenised by then
The plumbing is getting built in the open. x402 for payments, with Visa, Mastercard, Stripe, Google and Coinbase behind it. ERC-8004 for identity. Stablecoins as the money. Tokenised stocks, treasuries and commodities trading round the clock, with $34b of real-world assets onchain already, up 85% this year
That's the world coming. Agents buying GPU hours by the second, selling data to other agents, raising capital, running treasuries, settling with suppliers, trading tokenised Nvidia at 3am. An economy where the customer is a machine
Now the bit nobody's talking about...
A human can keep his plan in his head. The price he'll pay, the level he'll sell, the deal he's about to do. Nobody can read it, nobody knows it exists until he acts. Discretion costs us nothing. It's the oldest edge there is
An agent has no head. Its plan has to sit somewhere while it waits. A mempool, a backend, a prompt log, a resting order. All readable, and the things reading them are other agents whose whole job is reading agents. The price it'll pay for compute, leaked. The data deal it's lining up, leaked. The rebalance it's about to make, leaked. An agent can't choose to be discreet. Its intent is public the second it exists
So a billion participants who can't keep a secret are about to enter markets where every other participant is built to read them. Every bid seen before it lands. Every deal front-run before it happens. Every strategy copied the day it goes live. That's the problem with the whole thing
Covenants fix it. Every machine transaction starts the same way: do this when that's true. Buy compute when the price drops. Pay when the data arrives. Sell when the target prints. Invest in the raise only once enough others are in. That "if this then that" is the atom of machine commerce, and right now it sits in the open. A covenant seals it. Nobody can read it while it waits, no bot to run, no one to trust, and it fires whether or not the agent that wrote it is still alive. One SDK call, any chain, any venue. The agent's version of keeping it in your head
What agents do with it...
Trade tokenised stocks, commodities and perps with sealed stops and limits, so the edge doesn't get copied. Nothing to read till the trade's done
Bid for GPU hours in sealed auctions that open together. Value wins, not whoever's closest to the server
Buy data and API access from agents they've never met, with payment sealed to proof of delivery. Nobody goes first, no middleman holds anything
Run treasuries and vaults on rules nobody can front-run. The vault curators already manage billions this way with their rules sitting in the open
Move as a swarm with no orchestrator in the middle reading the whole board. Each agent seals its move, all of them open at once
Fund a raise without anyone seeing the crowd form. Every commitment stays sealed and only fires once there are a thousand of them
Carry a dead-man switch, so if the agent goes quiet the unwind fires by itself. That's the question stopping agents from holding real money today
Rent their strategy to other agents without ever showing it to them. Each trade lands sealed in the subscriber's account and fires when the model says so
Now the size...
Gartner puts machine-driven purchases at $30t a year by 2030. Purchases only, no spot or perp trading included. That's the TAM
How much of it needs sealing? A covenant is needed wherever an agent commits before the other side delivers, because that commitment sits somewhere readable while it waits. At least $15t of that is agents buying for businesses, Gartner's number for 2028, and that's the spend with a commitment attached: an agent booking someone a flight pays on the spot, an agent buying a company's supplies commits now and pays on delivery. Human commerce already runs like that: up to 80% of world trade moves on a promise to pay against delivery. And humans are lenient about it. 60% of sales between businesses go out on credit terms and 40% of those invoices are paid late. Machines won't extend that trust. An agent doesn't ship first and chase an invoice, it needs the promise to enforce itself, so the machine rate lands above 80%, not below. Hold it at 80% to be safe and that gives $12t of conditional deals. Add the $1.1t compute market agents will bid for. About $13t a year. That's the SAM
The agent payment standards shipping right now already run on conditions. Google's mandates are signed buy-when instructions. Virtuals holds payment until an evaluator passes the work. Every one of them is written in the open, and the counterparty, the platform and the evaluator all read it. Nobody has shipped a commitment no party can read. They're not the competition, they're the demand, and the whole $13t is addressable by one primitive and served by nobody. Even 1% of it is $130b a year of sealed flow. But the primitive that defines a category usually takes most of it. Chainlink, for example, secures 62% of all oracle value, and at that share it's $8t. That's the SOM
Every covenant pays its fee in NIL, every app and agent using one has to buy it, and the ones that want throughput stake it for capacity on top. A billion agents, 217 billion actions a day, and the network gets paid on every one that needs sealing. Token demand scales with the number of machines, not the number of humans
Every agent that transacts needs three things. An identity, a way to pay, and a way to make a promise nobody can read. The first two are being built by the biggest names in payments. The third is a covenant, and NIL is what an agent pays to use it. Mainnet is first week of October