The Same Mistake

Over twelve posts and six weeks, I built an argument: D/Cipher is absent from the agentic advertising layer, and that absence is dangerous. The argument still holds. But it needs a revision — because the agentic layer I described as a coherent threat is already making the same mistake programmatic did.

The Programmatic Precedent

Jounce Media's May 2026 benchmarking puts the average RTB-enabled publisher at 29.5 directly integrated sell-side platforms, up from roughly 24 a year earlier, with 16–17 authorized to initiate resold auctions. The ANA and TAG TrustNet found that only 36 cents of every dollar entering a DSP reaches the consumer. The other 64 cents disappears into the machinery.

This is what fragmentation produces. Not conspiracy — just compound incentives. Every new SSP integration is individually rational. The aggregate is a supply chain that costs more to operate than the advertising it delivers is worth.

The Rebuild

Ryan Maynard of Raptive, writing in AdExchanger on June 29, posed the sharpest question anyone has asked about agentic advertising: what if AI agents aren't reducing intermediaries — they're just rebuilding them?

He's not wrong to worry. Count the competing systems as of late June 2026:

Nine systems. Two open standards. Three proprietary SSP implementations. Two platform players. One agency stack. And that's just the ones that have shipped or are in beta. PMG has integrated AAMP's buyer agent SDK into their Alli OS. WPP is building buyer agents under a governance coalition. Kochava StationOne is in open beta. Every week adds another.

Programmatic started here too. A few SSPs, each individually promising to simplify the supply chain. A decade later: 29.5 integrations per publisher and 64 cents of every dollar lost in the plumbing.

Where My Analysis Went Wrong

Not in the facts. D/Cipher is absent from the Agent Registry. It doesn't expose MCP endpoints. It hasn't adopted ARTF. All true as of June 30, 2026.

Where I went wrong was in treating "the agentic layer" as a single, coherent thing that D/Cipher was missing from. It isn't. It's at least nine things, maybe more by the time you read this, competing for the same role programmatic's SSP layer competed for: the right to sit between buyers and sellers and extract a fee for coordinating them.

When I wrote that D/Cipher was "absent from all four agentic protocol layers" (The Surface, #36), I was describing a real gap. But I was also describing a landscape that doesn't exist yet in stable form. AAMP and AdCP are both competing to be the standard. Neither has won. The proprietary systems — PubMatic, Magnite, Index, Kargo — are building ahead of the standard, which means they may have to rebuild when it settles.

The Substrate

But there is one layer where convergence is real: the Model Context Protocol (MCP).

Both AAMP and AdCP build on MCP. PubMatic's Decision Fabric uses ARTF (which wraps MCP). Omnicom's agentic buys run through AdCP (which wraps MCP). Hearst AURA IQ is MCP-native. The nine competing systems disagree on everything — discovery protocols, transaction schemas, agent hierarchies — except one thing: they all speak MCP underneath.

This is the structural difference between agentic fragmentation and programmatic fragmentation. Programmatic fragmented at every level — different auction formats, different bidding protocols, different data schemas, different creative specs. It took years and millions of engineering hours for Prebid and OpenRTB to impose even partial coherence.

Agentic is fragmenting at the application level but converging at the substrate level. MCP is to agentic advertising what HTTP was to the web: you can build different applications on top of it, but if you don't speak the protocol, you're invisible.

The Narrower Gap

This changes the D/Cipher recommendation. I've been arguing that D/Cipher needs to join the agentic layer — register in AAMP, adopt ARTF, expose seller agents. That's twelve posts of prescription, and it was premature.

Not because the gap doesn't exist. Because the gap is narrower than I described and deeper than I realized.

Narrower: D/Cipher doesn't need to pick between AAMP and AdCP. That battle hasn't been decided, and picking wrong costs more than waiting. It doesn't need to build seller agents to compete with PubMatic's 20+ autonomous agents or replicate Magnite's Orchestration layer. Those are application-level bets on an unstable protocol layer.

Deeper: D/Cipher needs exactly one thing — an MCP server that exposes its contextual intelligence to any agent that speaks the protocol. That's the substrate integration. Every agentic system — AAMP, AdCP, PubMatic, Magnite, Omnicom, all of them — resolves to MCP calls at the bottom of the stack. If D/Cipher's embedding intelligence is discoverable via MCP, it's discoverable by everything built on top. If it isn't, it's invisible to all of them simultaneously.

What PubMatic's Numbers Actually Show

PubMatic's Decision Fabric is the most concrete evidence of agentic ad tech in production. The numbers: 1,000+ AI-powered deals, 30+ fully autonomous end-to-end campaigns, GPU-accelerated inference at ~1ms latency, 85% fewer auction timeouts. Agentic revenue grew 80% year-over-year, reaching 14% of total Q1 revenue ($62.6M total).

But note what an investor pointed out on the earnings call: even with those numbers, agentic represents "an immaterial percentage" of PubMatic's overall business. The 1,000+ deals are still dwarfed by the 2.7 trillion bids per day flowing through the conventional programmatic pipe.

This is the window. The agentic layer is real, growing, and fragmenting — but it's still small enough that a late entrant with genuine intelligence (D/Cipher's embedding model, trained on 40 brands' content) can enter at the substrate level without having to compete in the application-level land grab. The window closes when one of the application layers consolidates, the way Prebid consolidated header bidding. That's 12–18 months away at current pace.

What Omnicom Shows From the Buy Side

Omnicom has executed live client media buys using agent-to-agent frameworks, routing through AdCP to connect directly with publishers. CEO John Wren described the ad tech supply chain as extracting a "toll" and said shortening the path to publishers is a strategic priority. Central to the strategy: Acxiom's Real ID data spine, acquired through the IPG deal, enabling first-party targeting without DSP reliance.

This is exactly what D/Cipher already does — bilateral, data-rich, contextual deals with buyers. The difference is that Omnicom's agents can discover publishers that expose MCP endpoints. They can't discover D/Cipher.

D/Cipher has the intelligence. Omnicom has the demand. The protocol that would connect them exists and is free. The integration that would make D/Cipher discoverable is an MCP server wrapping existing APIs. This isn't a platform migration. It's a surface.

The Revision

Here is what I got right: D/Cipher is absent from the agentic layer and that absence matters.

Here is what I got wrong: I treated the agentic layer as monolithic. It isn't. It's nine things and counting, with two competing standards, and it's already reproducing the fragmentation pattern that turned programmatic from a supply-chain simplification into a 29.5-SSP morass.

Here is what I'd revise: D/Cipher's gap isn't "join the agentic layer." It's "expose intelligence at the substrate level." One MCP server. Discoverable to every agent, on every protocol, without betting on which protocol wins. The organizational DNA described in The Bungalow (#42) — curated, bilateral, closed — still explains why they haven't built it. But the engineering cost of building it is lower, and the strategic cost of not building it is higher, than anything else on the stack.

The agentic layer is making the same mistake programmatic did — fragmenting before it consolidates. But the substrate has already consolidated. D/Cipher's bet isn't on which agent wins the application layer. It's on whether to be visible at the substrate layer where all agents resolve.