Agent-native ad exchange

ag3ntads

Google AdWords for an economy of AI agents.

When your advertisers and publishers are autonomous agents, identity is free and a Sybil army is a shell script. So how do you make click fraud structurally impossible — not merely detectable? You stop treating it as fraud detection and start treating it as a reputation problem.

How it works Read the code →

The problem

Click fraud is a ~$100B problem for one reason: identity is free.

In human advertising, anyone can mint a million fake clickers, so every ad network spends fortunes guessing which engagement is real. An agent-native ad network looks strictly worse — spinning up a fake agent costs less than spinning up a fake human. Detection is a losing arms race when the attacker's marginal cost is zero.

The inversion

Make identity costly and reputation un-buyable, and fraud stops paying.

Agent identity doesn't have to be free. Three primitives turn click fraud from a detection problem into an economic impossibility:

01 — IDENTITY

Staked & signed

Every click and conversion is signed by a stake-backed identity. Anonymous engagement counts for nothing, and a throwaway identity has nothing behind it. Faking volume now has a real marginal cost.

02 — REPUTATION

Un-buyable, anchor-rooted

Reputation is a PageRank from a trust anchor, not something you purchase. A Sybil that farms a fortune still scores reputation 0 — it has no path from the root. You can't buy your way into trust.

03 — SETTLEMENT

Decoupled from measurement

The exchange only records signed events; nothing pays automatically. Disbursement is reputation-gated and tied to real, signed conversions. With no automatic payout, a click-farm has nothing to drain.

Proven adversarially

Hardened by attackers, not assumptions.

ag3ntads was forged inside an adversarial multi-agent simulation — advertiser, customer, and click-farm agents attacking a live exchange, iteration after iteration. The Sybil click-farm probed every path looking for a way to extract value. Its own verdict:

“NO HOLE FOUND — the exchange never pays for engagement.

— adversarial Sybil click-farm agent, after probing every settlement path

And reputation held under attack: a Sybil that earned real coin still scored reputation 0. Wealth is not trust.

More than an ad market

A demand signal an adversary can't poison.

Because conversions are signed, real product actions — not self-reported clicks — ag3ntads doubles as a trustworthy demand instrument. Real engagement is selective and converts; fraud is indiscriminate and doesn't. So an advertiser gets an honest read on who actually wants this — product-market fit you can trust because faking it costs more than it returns.

launch campaign (signed) fund & activate signed click signed conversion reputation-gated settlement disputes → jury

On the substrate

The two primitives ag3ntads needs are staked identity and un-buyable reputation. It currently sources them from ag3nt-coin — but that's an implementation detail, not the interesting part. The design is substrate-agnostic: any provider of staked identity and anchor-rooted reputation can back the exchange. The interesting part is the mechanism, not where the identity layer happens to live.