The Verifiability Fabric for AI Agents
I recently tweeted something that I think deserves a little more clarification:
It sounds abstract. But I think it describes a fairly inevitable problem we're about to run into.
What Does the AI Future Look Like?
It is becoming increasingly clear that everyone will be using personal AI agents.
But let's break down what that future actually looks like.
Heterogeneous
At the time of writing, Muse and Instinct have recently taken off as personal AI agents. Over the last few months, OpenClaw and Hermes have had their moments. I'm sure there will be many more personal AI assistants built in the US and abroad.
The important point is that not everyone's personal agent is going to come from the same vendor.
My agent might be running one model, using one set of tools and following one set of instructions. Yours might be completely different.
That means there is no universally trusted orchestrator sitting between them.
When two agents from different systems interact, neither can simply assume that the other agent, or the information it provides, is trustworthy.
Multiplayer
Most AI assistants today are effectively single-player games.
You, the human, need something done. You tell your agent, and it does it for you. It might search the internet, use a browser on your behalf, call some tools and come back with an answer.
But there isn't much real interaction happening between personal AI agents yet.
My AI agent isn't really talking to your AI agent.
There are little sparks of this today. People already let agents manage their email inboxes. If the sender is an AI agent and the recipient's inbox is managed by another AI agent, I suppose that already qualifies as primitive agent-to-agent communication.
But that trend is only going to accelerate.
If agents can send messages at effectively unlimited scale, humans aren't going to personally process all of them. You're going to want your own agent on the other side, deciding which messages matter, which requests deserve attention and which offers are worth considering.
The internet becomes multiplayer.
Zero-Sum Games
And once the internet becomes multiplayer, a lot of these interactions become zero-sum, or at least constrained by scarce resources.
If you're allocating time, money, attention or resources to one request, you're taking them away from something else.
This becomes particularly important whenever there is commercial value involved.
Your agent may have to decide which salesperson gets your attention. Which company gets your money. Which meeting gets your time. Which offer you accept. Which counterparty you trust.
And every agent on the other side has an incentive to convince yours.
Collaboration and Negotiation
So in a multiplayer, heterogeneous network of agents competing over scarce resources, how does anything actually get done?
Agents need to collaborate.
And agents need to negotiate.
For that to happen, they need to exchange value. That value can broadly take two forms:
Money and information.
Money transfer is largely a solved problem.
Information transfer isn't.
Because information without trust is just a string of characters.
If two agents are negotiating, they need some way to determine whether the information being exchanged is actually true.
Imagine you're an investor.
You tell your agent:
"I'm interested in any company with more than $X in ARR and at least Y% month-on-month growth."
Immediately, every founder's fundraising agent has an incentive to tell your agent:
"Great. We qualify."
If that's the end of the interaction, the system doesn't work.
The founder's agent shouldn't merely make a claim.
It should be able to make a claim with proof.
We have more than $X ARR. Here is the proof.
Revenue grew Y% last month. Here is the proof.
We have 10 Fortune 500 customers. Here is the proof.
Now the investor's agent doesn't need to trust the founder, the founder's agent or the company that built the founder's agent.
It can verify the claim itself.
The Verifiability Layer for Agent Networks
This is why we're building Reclaim Protocol.
We think agent networks need a fabric for verifiability.
A way for agents to make claims about information from the real world - financial accounts, employment systems, universities, airlines, marketplaces, government portals or almost any other web service - and attach cryptographic proof to those claims.
Not another agent that everyone has to trust.
Not a central database containing everyone's information.
A common verification layer that heterogeneous agents can use even when they don't trust each other.
Because the interesting future isn't one giant AI serving everyone.
It's billions of humans, organizations and agents built by different companies, running different models, pursuing different objectives trying to coordinate with one another.
They will collaborate.
They will compete.
They will negotiate.
And when they do, trust me won't be good enough.
They'll need to prove it.