A founder's note · Brian O'Kelley
This has happened before.
I built the first ad exchange.
I know exactly what it got right. And I know what it broke.
In 2003, at Right Media, I wrote the code that connected separate pools of advertising demand and supply into a single market. The result was counterintuitive: advertisers paid less, publishers earned more, and the network made more money. All at the same time.
It was the first time I understood what an exchange actually does. It doesn't split value. It creates it. By connecting liquidity pools that were previously isolated, the exchange expanded the market itself. That insight became real-time bidding. RTB became programmatic. Programmatic became a $600 billion industry.
I went on to co-found AppNexus and spent a decade scaling that infrastructure. I've spent twenty years inside the system that grew from that original code.
What went wrong.
Programmatic advertising funded the open web. It put billions of dollars into the hands of content creators. That part worked.
But the system we built also created something ugly.
A single ad impression on a major news site now triggers 55 network calls, exposing reader data to five separate companies before the page finishes loading. Data isn't the new oil. Data is fallout, an uncontrolled byproduct of a system that was never designed to contain it.
Between the brand's dollar and the publisher's revenue, there are fifteen intermediaries. Each takes a cut. Each adds latency. Each adds opacity. By the time a dollar reaches a publisher, roughly half has been absorbed by the supply chain.
The internet didn't lose to walled gardens because it was less valuable. It lost because it was harder to buy.
The incentives are adversarial. DSPs optimize for buyer outcomes. SSPs optimize for seller yield. Everyone in between optimizes for their own margin. Nobody optimizes for the system.
None of this happened because the technology was wrong. What went wrong was the accumulation of intermediaries, the misalignment of incentives, and the erosion of trust between the parties the system was supposed to connect.
You can't fix this by adding another layer to the existing stack. You can't rebrand a horse as an autonomous vehicle and hope it survives on the highway.
You have to rebuild.
The shift.
Here's the conceptual leap that makes the rebuild possible.
Programmatic advertising is a valuation system. Its core question is: What is this impression worth? An auction runs. A price is discovered. A transaction executes. Billions of times a day.
That was the right question for the era of human-operated dashboards and manual campaign management.
It's the wrong question for agents.
An AI agent doesn't need to value individual impressions. It needs to allocate a budget. The question changes from "what is this impression worth?" to "given my objectives, my constraints, and every available opportunity across every channel, how should I distribute this spend?"
Programmatic asks
What is this impression worth?
Agentic asks
How should I allocate this budget?
That's not an incremental change. It's a different discipline.
This is the difference between high-frequency trading and portfolio management. RTB is the former. What agents need is the latter.
Most advertisers today buy from three to five platforms, because the execution cost of managing each relationship scales linearly with complexity. A human can only operate so many interfaces. That constraint has nothing to do with strategy. It's a transaction cost problem.
Agents eliminate it. When the cost of managing a supplier relationship approaches zero, an advertiser can diversify across twenty or fifty or a hundred sources. Portfolio theory does the rest.
This is the same insight from 2003, that connecting isolated liquidity pools creates value for everyone, now applied at a completely different scale.
The programmatic model
- Brief
- Campaign Manager−15%
- DSP−10%
- RTB Auction−5%
- SSP−10%
- Data & Verification−11%
Publisher receives ~$0.49
The agentic model
- Brief
- Buyer agent discovers sellers
- Multi-turn negotiation
- Governance validates
- Platform fee−4–8%
Publisher receives $0.92–0.96
The shift is structural.
View pricing
See pricingWhat an agentic ad platform actually is.
It's a marketplace built from the ground up for machine-to-machine commerce.
-
1. Discovery
A buyer agent receives a brief: "Reach environmentally conscious millennials in the Northeast, $50K budget, mix of video and audio." The agent queries the platform to find sellers whose inventory matches, across CTV, podcasts, social, display, out-of-home, whatever surfaces are relevant. Discovery is not a keyword search. It's a multi-dimensional match.
-
2. Negotiation
The buyer agent and seller agents negotiate directly. They iterate on availability, pricing, targeting, and terms across multiple turns. This is closer to how media has always been bought in direct sales, at machine speed, at any scale, for any deal size.
-
3. Governance
Before execution, a governance agent validates the transaction. Does it comply with the brand's safety standards? Does it meet sustainability requirements? Is the supply path clean? Governance isn't a report you pull after the campaign runs. It's a gate the transaction passes through before a dollar moves.
-
4. Execution
The deal executes. Creative is delivered. Impressions are served. Performance data flows back to the buyer agent, which uses it to refine the next negotiation. Every step is logged. Every fee is visible.
-
5. Trust
A file called
adagents.json, analogous torobots.txtorads.txt, sits on each publisher's domain and declares which agents are authorized to transact on their behalf. Buyer agents verify this before engaging. Bidirectional trust, built into the protocol.
When an AI agent can discover your inventory, negotiate terms, and execute a buy without human involvement on either side, the friction cost approaches zero. That changes everything.
A concrete example.
A sustainable fashion brand wants to sponsor a climate podcast. Budget: $500.
Under the current system, this deal doesn't happen. The sales cost exceeds the purchase margin. No human sales rep is going to negotiate a $500 sponsorship. No media buyer is going to spend two hours in a platform setting it up.
With an agentic platform, the buyer agent discovers the podcast through the seller's storefront, negotiates a host-read spot, validates brand safety, and executes. In seconds, for pennies of transaction cost.
The podcast gets funded. The brand reaches the right audience. The deal that economics previously made impossible becomes trivial.
The economics.
The original exchange insight was economic, not technological. Connecting separate liquidity pools didn't just make trading more convenient. It expanded the total market. Everyone was better off.
The agentic platform does the same thing. But it also collapses transaction costs in a way the original exchange never could.
Transaction costs today
A brand wants to buy across CTV, social, podcasts, and display. Four separate platforms. Four logins. Four optimization workflows. Four reporting systems.
The supply chain between the brand's budget and the publisher's revenue contains multiple intermediaries, each taking 5–15%.
Cumulative take rate: ~50%
Transaction costs with agents
The buyer agent discovers sellers across all four channels through one protocol. Negotiates, executes, and optimizes simultaneously.
Transparent platform fee. No hidden intermediaries. No daisy chains.
Transparent platform fee: 4–8%
When transaction costs drop from 50% to single digits, the economics of every deal change.
Deals that were too small become profitable. A $500 podcast sponsorship. A $2,000 local news campaign. A $200 out-of-home placement during a community event. These transactions were always worth doing. They just weren't worth the overhead.
Imagine an internet where ad tech makes publisher revenues go up, not down. That's not idealism. It's arithmetic.
When the long tail of content (local news, niche podcasts, minority-language publishers, independent creators) becomes economically viable to buy, the entire market grows.
Ready to explore more?
Governance as architecture.
Every ad tech company claims transparency. Most of them mean they'll send you a report.
We mean something different. We mean governance is in the protocol. Not bolted on after the fact, not available as a premium add-on, not dependent on a vendor choosing to participate.
Orchestrator
Proposes the transaction. Defines objectives, constraints, and budget. Cannot validate its own proposals.
Governance agent
Independently validates against brand safety, supply path standards, sustainability requirements, and civic policy. Approves, flags, or blocks.
Seller
Fulfills the validated transaction. Delivers inventory, creative placement, and reporting. Cannot override governance.
Separation of concerns. The proposer never validates. The validator never fulfills. Trust is structural.
Progressive autonomy.
Three modes: Audit (log every decision), Advisory (flag concerns before execution), Enforcement (block non-compliant transactions). Move along the spectrum at your own pace.
Every decision, logged.
When an agent buys media, the governance record captures the brief, inventory evaluated, why a seller was selected, what checks were performed, and the outcome. Full accountability.
Brand safety in the transaction.
Standards are encoded in the governance rules that every transaction passes through before it executes. Agents can't accidentally violate policy.
Governance cannot be a premium feature. It has to be the default.
The civic case.
Advertising infrastructure is civic infrastructure. This is not a metaphor.
What gets funded shapes what gets published. What gets published shapes what gets read. What gets read shapes what people believe is true.
When the advertising system makes it expensive to buy ads on local news, local news loses funding. When local news loses funding, it shuts down. When it shuts down, communities lose the information they need to govern themselves.
We've watched this happen for fifteen years.
Agents change this. When transaction costs approach zero, the economics of advertising no longer punish small publishers. A buyer agent can just as easily transact with a local newspaper in Cleveland as with a national platform.
Local journalism becomes viable.
Not through charity or public subsidy, but through advertising economics that actually work for local audiences.
Minority-language content gets funded.
When the cost of buying an ad on a Spanish-language podcast or Mandarin-language newsletter is the same as buying on an English-language platform, advertisers follow the audience.
Public interest advertising scales.
Voter registration campaigns, public health information, emergency services: all become dramatically cheaper to execute when agents handle the buying.
Civic campaigns reach communities.
A public health campaign doesn't need a $500K minimum. An agent can assemble a media plan across local radio, community newsletters, and regional social. In minutes.
Advertising infrastructure is not a neutral system. It is a civic system.
The protocol.
The Advertising Context Protocol, AdCP, is the open standard that makes all of this interoperable. Built on the Model Context Protocol (MCP). Open source. Apache 2.0 license. Over twenty companies collaborated on version one.
AdCP covers eight domains: media buying, creative, signals, accounts, governance, brand identity, sponsored intelligence (advertising inside AI products), and trusted match (privacy-preserving identity). Any buyer agent that speaks AdCP can transact with any seller agent that speaks AdCP, across social, CTV, search, podcasts, out-of-home, influencer, retail media, and AI services.
The protocol is surface-agnostic by design. When a new medium emerges, agents can buy it on day one. The transaction primitives are universal.
The invitation.
Twenty years ago, a small team built the technology that became the ad exchange. The architectural decisions made in that codebase shaped how hundreds of billions of dollars moved through the advertising economy for two decades.
Some of those decisions were right. Some created problems that the industry is still trying to solve. But all of them were made in a narrow window, by a small number of people, before anyone fully understood the consequences.
We're in that window again.
The decisions being made right now, in protocol specifications, in working group sessions, in GitHub repositories, will determine how advertising works for the next twenty years. Whether agents transact through open protocols or proprietary APIs. Whether governance is structural or optional. Whether the economics favor the internet or accelerate its decline.
The internet will not be saved by better technology. It will be saved — if it is saved — by people making the right architectural decisions. That moment is now.
— Brian