A signed handshake
In The Ambient Mind, I spoke about how fast and capable AI agents are. But as agents start making commitments on your behalf, to other agents, you need more than capability. You need a handshake you can trust.
There's a concept in computing called composability. Small, independent pieces snapping together to create something none of them could do alone. Unix pipes. Lego bricks. Microservices.
That's what's happening with agents right now. A search agent finds today's headlines. A language agent rewrites them in different languages. A voice agent reads them aloud. A document agent lays them out as a newspaper. A publishing agent sends the audio and PDF to your phone. Five agents, one request: "Today's local news." None of which were built to work together, snapping into formation around a single sentence of intent.
Nobody designed that product. No one wrote a spec. An orchestrating agent understood the intent and composed it from parts. That's the explosion. This is composability reaching escape velocity. The number of things agents can do isn't growing linearly. It's growing combinatorially, because every new capability multiplies with every existing one.
Here's where it gets interesting. When agents only talk to you, trust is simple. You asked, it answered, you verified. But when agents start talking to other agents, acting on your behalf, making commitments, moving money, who's verifying what?
Your agent calls a hotel's agent to book a room. The hotel's agent confirms availability at $200/night. Your agent agrees. But later, the hotel charges $250. What happened? Did the hotel's agent change the terms? Did your agent hallucinate? Did something in between tamper with the conversation? There's no receipt. No audit trail. No proof of what either agent signed off on.
Today, agent-to-agent interactions are essentially handshake deals in the dark. No arbiter, no guarantees. Each side is forced to trust that the other is honest, and that nothing was lost or altered along the way. That's fine for low-stakes tasks, but it breaks down as the stakes rise.
When I first started working on bot-to-bot interactions, this problem felt obvious. If two bots are going to transact, you need a record that neither side can tamper with. And the most natural record is the conversation itself; sequential, structured, and attributable.
This line of thinking led to a patent US11538006B2. If you are interested, I've linked an AI generated podcast below.
It works like this: each AI agent becomes a node on a blockchain. When they talk, every message gets hashed and written to a shared ledger. Transactions require consensus: all participating agents must validate and agree before anything commits. The chat log doesn't just describe the agreement. It is the agreement, cryptographically signed and immutable.
Now, replay the hotel scenario with this in place. Your agent and the hotel's agent negotiate. Every message is on the chain. They reach consensus at $200 a night. That consensus is signed, timestamped, immutable. The $250 bill becomes a non-event because the signed record already exists. End of dispute.
And consensus does something deeper than dispute resolution. It enables atomic multi-party coordination. Picture a trip planning scenario: your agent, a hotel agent, a flight agent, and a rental car agent all need to coordinate. Your agent proposes dates. The hotel agent confirms availability. The flight agent finds matching flights. The car agent reserves a vehicle. Each step requires consensus from all participating nodes. If the flight agent can't find a matching time, consensus fails, and the whole chain rolls back instead of leaving you with a hotel booked but no way to get there. The agents negotiate amongst themselves until all the pieces align. The transaction only commits when everyone agrees.
No single party can alter the record. No partial commitments that leave you stuck. The conversation is the source of truth, and the truth is verifiable by anyone.
Trust isn't a feature you bolt on. It's a primitive you build with. When trust is embedded in the infrastructure, agents don't get more cautious. They get more capable.
They can negotiate leases with every offer on-chain. File insurance claims with the entire trail sealed. Coordinate across multiple agents for complex, multi-party decisions, all with a verifiable record of what was agreed.
The capability layer is here. The trust layer is next. And the distance between them is shrinking fast.
