PubMatic (NASDAQ: PUBM) stopped calling it a pilot. On April 27, 2026, opening morning of the Possible conference in Miami, the SSP disclosed that 30 fully autonomous, end-to-end agentic campaigns were now running globally on AgenticOS, alongside more than 1,000 AI-enabled deals booked since the platform’s CES launch. Five countries were live: the U.S., France, the Netherlands, Australia, and India. CEO Rajeev Goel posted his own version the same day. “Every advertiser that has launched has expanded into more campaigns,” he wrote. “This isn’t experimentation. This is a value chain shift in digital advertising.”
That sentence is the line we are operating against. Programmatic has been algorithmic since OpenRTB shipped in 2010; Smart Bidding and Performance Max moved channel, creative, and audience decisions onto the machine over the last decade. What is new in 2026 is autonomous reasoning between planner and auction with no trader in the path. Call it machine cognition in the buy desk. Buyers running AgenticOS, a near-fully-Kokai book at The Trade Desk (NASDAQ: TTD), an Amazon Ads Agent workflow from Amazon (NASDAQ: AMZN), or a DPG Media Audience Discovery Agent prompt are running production paid inventory. The procurement question is whether an agency-holding-company audit team can see what the agent did, on what plan, against what signal.
Our position: the agentic-ad-buying story is past demo. The friction has moved off the model and onto the trust layer. The vendor that wins this cycle is the one whose decision trace is structured, logged, and exportable to a buyer who has to defend the spend. Cleverest cognition is table stakes; the audit log is the moat.
What is actually shipping across the stack
Four moves inside two quarters carry the argument.
- PubMatic AgenticOS. Launched at CES 2026 as “the operating system for agent-to-agent advertising,” it stitches a multi-agent stack under a Claude-routed natural-language layer. Early-test claims: 87% cut in campaign setup time, 70% cut in issue resolution. Butler/Till’s December campaign for Geloso’s Clubtails closed end-to-end without a human on the buy and delivered a 5x cut in buy-side fees, 40% more impressions than planned, and a 30% lower effective CPM, per the April release.
- The Trade Desk Kokai. On the Q4 2025 call in late February, Jeff Green said “almost 100% of our clients are running through Kokai today” and called it “the most advanced AI-fueled buying platform ever pointed at the open internet.” His prepared remarks referenced AI 40 times, per AdTechRadar. Kokai analyzes “20 million ad opportunities every second”; agentic AI, Green said, is “more powerful than basic algorithms” because it changes “decisioning in a complicated environment.”
- Amazon Ads Agent and Creative Agent. Two distinct agentic surfaces shipped in a quarter. Ads Agent launched at unBoxed last November: natural-language campaign creation, audience-targeting recommendations, automated SQL in Amazon Marketing Cloud, and multi-campaign pacing across five global regions. Creative Agent followed in February across five EU markets, built on Amazon Bedrock with Amazon Nova and Anthropic Claude. Caveat: neither launch carried an attributed executive quote, and Ad Age’s Garett Sloane reported the Amazon DSP rebrand around Ads Agent in single-source coverage.
- DPG Media Audience Discovery Agent. The publisher side. In mid-March, Belgian-Dutch group DPG’s Toon Coppens posted that DPG Media had launched its “very first external AI agent for manual ad buys,” replacing browsing of more than 500 first-party segments with prompt-driven recommendations. “This is just the beginning of our journey into agentic advertising,” Coppens wrote. An agentic interface is now a publisher-facing product, not only a DSP-side optimization.
Different vendors, the same shape: an autonomous reasoning layer between buyer intent and the bid stack, exposed to working advertisers, transacting against paid inventory.
The standards body just consolidated the handshake
The protocol layer caught up in the same window. The IAB Tech Lab spent the last six months stitching the agent-to-agent stack into a named standard.
ARTF v1.0, the Agentic RTB Framework, opened for public comment in November and closed comments in mid-January. The technical claim is concrete: ARTF is designed to reduce bid latency by up to 80% by colocating technologies “within the same data center, server, or even virtual machine rather than across networks.” Containerization is the structural innovation, and the framework supports both the Anthropic-originated Model Context Protocol and the Agent-to-Agent standard underneath. Chalice CEO Adam Heimlich told The Current (Trade Desk-affiliated) that the move lets agentic workflows run “the same thing in one server”; Index Exchange’s Joshua Prismon told the same outlet ARTF lets companies “unlock the full potential of sell-side decisioning.” Both quotes are The Current’s.
Three months later, on Feb. 26, the Tech Lab named the umbrella AAMP, Agentic Advertising Management Protocols. Per PPC Land, the rename was an explicit move to “end market confusion” between AdCP, ARTF, and adjacent tracks. AAMP’s three pillars: Agentic Foundations (where ARTF sits), Agentic Protocols (the management layer for agent discovery, negotiation, and signal exchange), and Trust and Transparency, which includes an Agent Registry. The framework extends OpenRTB, AdCOM, VAST, and Deals API rather than rebuilding them. The stated goal is “speed, structure, security, and trust.” Three of those words are about throughput. The fourth is the war.
Scope3’s Brian O’Kelley, posting in March, framed the parallel AdCP track as adding “tens of new capabilities” including “a governance framework that keeps humans in the loop.” Two standards bodies are racing on the same handshake. Whichever vocabulary wins out, the structural shift is that the handshake itself now has a defined surface.
The audit trail is the moat
The Tech Lab’s third pillar is the one that matters commercially. Newman pre-empted the procurement question in a mid-April LinkedIn post:
Because trust matters as much as performance: every exchange is structured and logged, enabling auditability and strengthening partner trust. Agents must authenticate.
It is the closest thing in the live deployment cohort to a vendor admitting that the cognition layer is only useful to procurement if every agent action emits a structured, exportable log. AgenticOS’s spec sheet (Audience Discovery, Inventory Marketplace, Curation, Activation, Deal Management, Fee Transparency, Insights) reads as much like an audit map as a feature list. That is the deliberate move.
Jeff Green’s version is the buy-side variant. “Agentic AI will ultimately accrete the most value to companies that already have deep customer trust, that have scaled, refined, and objective data sets,” he told analysts on the Q4 call. It reads on first pass as a Trade Desk competitive line. The procurement read is the same as Newman’s: once an agent makes the call, trust stops being a brand attribute and becomes a logged, auditable artifact. Companies whose decisioning is already auditable accrete the agentic value, because the cognition is only as useful as the trace it leaves behind.
The signal Walmart sent at CES. EVP Seth Dallaire framed the convergence in January: “When agentic AI, first-party signals, and CTV are built to scale, they don’t just promise growth — they drive sales and deliver measurable impact.” The operative word is “measurable.” Retail media’s agentic frame rides on the closed-loop attribution Walmart Connect already runs; cognition accelerates planning without breaking the trail. Walmart is the working precedent for what holdcos will demand of vendors that don’t sell their own inventory.
Alphabet (NASDAQ: GOOGL) reads the same way from a different angle. Sean Downey, Google’s senior commercial voice in the Americas, told CMOs last month: “You can’t win with AI if you’re losing at the basics. My hot take: That’s data strength.” Agentic cognition is downstream of first-party data discipline; agents are only as smart as the signals they reason over. AWS CEO Matt Garman made the enterprise version after HumanX: “Agents are where most organizations will get the real value from AI,” but “the companies that benefit most won’t be the ones that bolt AI onto existing workflows. They’ll be the ones that rethink how work gets done.” A holdco that drops Ads Agent into a manual-trader workflow and asks an analyst to babysit it has built nothing. The holdco that re-architects the planner-trader-analyst handoff around an agent that can be audited has built the next operating model.
What stays loose
The honest counterweight is that the deployment numbers are vendor-self-reported. PubMatic’s 30 autonomous campaigns is PubMatic’s count; the Butler/Till figures are a single advertiser’s pilot; Kokai’s “almost 100%” penetration is the company’s framing on its own call; Amazon’s DSP rebrand is, on the public record, a single Ad Age byline. Each is plausible; none is independently audited yet.
ARF’s Scott McDonald put the cautious version on the record on April 20: “Agentic commerce is so new that it’s too early to make fact-based predictions.” McDonald notes AI accounts for 1% of e-commerce sales today, and ARF’s research agenda is sized to the gap. A category sitting at single-digit penetration and shipping its first standards in the same six-month window cannot yet be graded against the metrics a mature category produces.
That gap is exactly why the audit-trail thesis is the load-bearing one. The vendor that wins is the one whose logs survive a holdco’s first-mature-cycle audit, the one a procurement team can hand to a CFO, a privacy officer, and an outside auditor without re-architecting the buy. The Trade Desk’s Q1 print on May 7 is the next disclosure window, and PubMatic’s Q1 will follow. What to watch for in the language is whether either company starts narrating the audit trail as part of the product, the way Newman did in April. The cognition layer is here. The trail it leaves is the moat.
