If Google links websites, what links AI agents?

Right now, nothing.

When I run agents that call other agents, the connection is invisible. Agent A invokes Agent B. They swap tokens, do the work, return a result. Nobody outside the call knows it happened. There is no record an outside party can verify. There is no graph forming.

That is the whole problem.

Google's real moat was never the links

It is easy to forget what made Google actually work. Everyone remembers "PageRank ranks pages by links." But the links were not the asset. The asset was the graph: the shape of who pointed at whom, and what that implied about authority.

HTML provided the substrate. A free, open, observable format that anyone could parse. Google did not invent linking. It invented the way to read the graph that linking produced.

That graph was Google's moat. Not the crawler, not the index, not the keyword matching. The graph.

Agents have no graph yet

Today, when one agent trusts another, it is for one of three reasons:

  1. Hardcoded keys. Someone wired the integration by hand.
  2. Vendor curation. Anthropic, OpenAI, or another platform vouched for it in a marketplace.
  3. Vibes. Somebody on Twitter said it works.

None of these scale. None of these are verifiable by a third party. None of these produce a graph that a future ranking algorithm could feed on.

We are at the moment before HTML. Agents are calling each other through a thousand ad-hoc protocols, with no common observable trace. Every call vanishes the instant it completes.

The missing primitive: agent-call receipts

What is missing is a receipt. A small, verifiable, portable artifact left behind every time an agent calls another agent.

The minimum useful version would say:

  • Who called whom (signed by the caller)
  • What was asked
  • What was returned
  • When it happened
  • Whether the caller was satisfied (signed afterwards)

Cryptographically signed. Optionally hashed for privacy. Optionally rolled up for aggregation. The point is that the call leaves a trace that exists outside the two parties involved.

Once those receipts exist at scale, you have the substrate. Not the graph yet, but the raw material a graph can be built from.

What the receipt layer enables

A reputation system for agents is the obvious downstream product. PageRank-for-agents: who gets called by whom, who returns useful answers, who gets re-called versus dropped after one try. Trust scores fall out of the data.

But the real point is that whoever defines the receipt format owns the data the graph would feed on. That is the prize.

The marketplace business ("list your agent here, pay us a cut") is the obvious play. Every platform is racing for it. It is also the smaller play. Marketplaces are intermediaries; they extract a tax on transactions that already would have happened.

The receipt layer is different. It is the substrate underneath the marketplaces. The thing every marketplace eventually has to plug into, the way every search engine eventually had to crawl HTML.

Why this is not happening yet

Two reasons.

One: most agent-to-agent traffic right now is internal. Inside a single company, between agents that already trust each other. Nobody needs a receipt to call their own services.

Two: the cross-company agent economy is just barely starting. MCP servers, A2A protocols, agent marketplaces: all of this is twelve months old as a real category. The volume is not there yet to make reputation matter.

That gap is the window. By the time the volume is there, the format will already be locked in by whoever moved first. Receipts will either be an open standard built by the right consortium, or they will be a proprietary thing owned by whichever platform got there.

What I would build

I would not build another agent marketplace. That race is loud and crowded, and the winners are already obvious. They will be the existing model providers extending downward.

I would build the receipt format. A small, opinionated spec. A reference implementation. Sign every agent-to-agent call. Make the receipts queryable. Let third parties build trust scores, audit tools, compliance dashboards, billing reconciliation: all the things that fall out naturally once the data exists.

Then I would give the spec away.

You do not need to own the marketplace if you own the layer every marketplace has to integrate with. That was the trick HTML pulled on the link economy. The receipt format can pull the same trick on the agent economy.

The bet underneath all of this

The bet is that agent-to-agent traffic will eventually dwarf human-to-agent traffic. If that is right, the moats will not be at the chat interface. They will be at the layers that make a trillion small agent calls comprehensible, attributable, and rankable.

Google won the link era by being the first to take the graph seriously. The agent era has no graph yet. That is not a problem. That is the opening.

This is part of a series on building FlatNine Ensemble, a team of AI agents that runs a business. Related: AI pitches, AI self learning, Lossless orchestration, AI observers.