The Two Paths of Agentic Media Buying
In the past month, agentic advertising went from conference talk to shipped product at nearly every major platform. But "agentic" now describes two fundamentally different architectures: AI agents bolted onto the existing programmatic and RTB stack, and AI-native infrastructure where agents are the transaction layer itself. For the buy side, this creates a problem that is easy to miss in the launch noise: budgets do not care which architecture wins. Outcomes live on both paths, and a buying platform that only speaks one of them will leave results on the table.
This post breaks down what happened in the past several weeks, defines the two paths precisely, and makes the case for why AI-native agentic media buying, the approach Adgentek built ORCA around, is the only buy-side architecture that delivers value across both. The short version: we are at the very beginning, both paths are running live campaigns, and the winning position on the demand side is not picking a path. It is orchestrating across them.
What happened in agentic advertising over the past month
Between late June and mid-July 2026, agentic ad infrastructure shipped from Amazon, Meta, Google, OpenAI, FOX, WPP, Yahoo, and Warner Bros. Discovery, marking the first wave of real product after a year of announcements. The concentration of launches around Cannes Lions in late June made one thing clear: the industry has moved past experimentation and is rebuilding its infrastructure layer for AI agents.
The specifics, in rough chronological order:
- FOX Advertising launched what it calls an end-to-end agentic advertising platform on June 17, powered by FOX AdStudio, embedding agents across audience planning, media transactions, and activation for its linear and digital inventory.
- WPP launched its Buyer Agent for Video, an AI planning agent that evaluates inventory across premium media owners including Disney, Netflix, NBCUniversal, Paramount, and Fox. Notably, the agent recommends and humans approve; WPP co-built the governance framework with IAB Tech Lab and Prebid.org.
- Yahoo DSP announced an Agent Network connecting advertisers to AI tools from 23 ad tech partners, extending its framework that lets advertisers bring their own agents, use Yahoo-built agents, or combine both over MCP and APIs.
- At Cannes Lions, Amazon debuted Alexa+ Agentic Ads in beta, a format where the consumer can transact inside the ad unit itself. Meta and Google shipped agentic creative and campaign tooling in the same window, and OpenAI attended Cannes for the first time, pitching ChatGPT as an ad channel and citing that roughly one in five queries carries commercial intent.
- OpenAI followed up in early July with job listings for engineers to build text, image, video, native, conversational, and interactive ad formats, signaling a move well beyond the static sponsored links it launched on February 9, 2026 at a $60 CPM.
- Warner Bros. Discovery announced in July that it is rebuilding its advertising stack on AWS with agentic AI decisioning across planning, activation, optimization, and monetization for its US linear and digital channels.
Against that product wave, the adoption data tells a humbler story. Boston Consulting Group's June survey of 300 CMOs found that while 96 percent say AI is transforming their marketing function, only 8 percent are running campaigns where multiple AI agents operate autonomously. A Digiday survey of 180 agencies, publishers, brands, and retailers found most describing the industry as either experimenting but fragmented or early but moving quickly. The infrastructure is arriving faster than the operating model. That is what the very beginning of a platform shift looks like.
Path one: agents layered on programmatic infrastructure
The first path keeps the existing programmatic machinery, OpenRTB auctions, DSPs, SSPs, and bidstream data, and puts AI agents on top of it as an automation and intelligence layer. This is where most of the announcements above live. The agent gathers campaign details, builds the media plan, traffics line items, diagnoses pacing issues, and reallocates budget. But when money actually moves, it moves the way it always has: through a bid request and response clearing in milliseconds across the traditional chain.
The Trade Desk's Koa Agents, PubMatic's AgenticOS, Yahoo DSP's agent framework, and the holding-company copilots all follow this pattern. So does IAB Tech Lab's AAMP (Agentic Advertising Management Protocols), which formalizes the approach at the standards level by extending OpenRTB, OpenDirect, and AdCOM into agent-operated workflows rather than replacing them.
This path has real strengths. It plugs into a trillion-dollar spend base, existing demand relationships, and measurement infrastructure that took two decades to build. Adoption is incremental: an agency can hand pacing diagnostics to an agent this quarter without changing anything else. But it also inherits every structural assumption of the stack underneath it. The transaction is still a sub-100-millisecond auction on an impression. The signal is still page context and behavioral inference. The fee stack is still the fee stack. An agent operating this machinery is faster and cheaper than a human trader, but it is trading the same asset on the same terms.
Path two: AI-native agentic infrastructure
The second path builds the advertising stack from scratch for a world where AI agents are participants in the transaction, not operators of the old one. Here, a buying agent and a selling agent communicate directly over protocols built for agents, primarily Model Context Protocol (MCP) and the Ad Context Protocol (AdCP), the open industry standard for agentic ad transactions, of which Adgentek is a Founding Member.
The architecture is fundamentally different from RTB. Where OpenRTB clears impression auctions in under 100 milliseconds, agent-to-agent transactions are conversational and asynchronous: a buying agent describes an objective, a selling agent responds with matching inventory, terms are negotiated, and the deal can be written directly into an ad server, often without touching the bidstream at all. FreeWheel proved the model in production in January 2026 with the first MCP-based agent-to-agent media transaction. Omnicom has since disclosed executing live agent-to-agent buys for clients using an AdCP-based framework, explicitly to shorten the supply chain and reduce intermediary fees.
This path also unlocks inventory the legacy stack structurally cannot reach: AI chatbots, AI search engines, AI assistants, and autonomous agents, where intent is expressed in natural language rather than inferred from behavior, and where formats are conversation-native rather than slot-based. That is the world of conversational advertising, served on the supply side by infrastructure like Adgentek's Agentic Ad Server. But the deeper point for buyers is that path two is not just new inventory. It is a new transaction model that works for any media a selling agent represents, from AI surfaces to CTV to direct publisher deals.
How the two paths compare
The clearest way to see the difference is side by side. The paths diverge on transaction layer, signal, inventory, and what the agent actually does.
| Dimension | Agentic on existing infrastructure | AI-native agentic |
|---|---|---|
| Transaction layer | OpenRTB bid request and response through DSPs and SSPs | Agent-to-agent over MCP and AdCP, often bypassing the bidstream |
| Timing model | Sub-100ms impression auctions | Asynchronous negotiation, from seconds to full deal cycles |
| Intent signal | Inferred from page context, device IDs, behavioral data | Expressed directly in natural language or structured task parameters |
| Inventory | Existing web, video, and CTV ad slots | AI surfaces and autonomous agents, plus direct agent-negotiated deals into traditional media |
| Role of the agent | Workflow automation operating legacy machinery | Transaction participant negotiating and executing deals |
| Fee structure | Traditional intermediary stack intact | Compressed chain with fewer intermediary layers |
| Governance model | Platform controls and human sign-off bolted onto tools | Judgment and policy enforcement built into the agent architecture itself |
Why the buy side should not pick a path
A media buyer's job is to deliver an outcome, not to defend an architecture, and in 2026 the outcomes are distributed across both paths. Most of today's reach still clears through programmatic pipes, and the efficiency gains from agent-operated workflows are real; that is why 70 percent of current agentic investment is going into media buying optimization. At the same time, the highest-signal new inventory, the 800M+ weekly AI assistant users and the agent-negotiated direct deals that FreeWheel and Omnicom have proven out, only exists on the AI-native path.
This is where the copilot approach hits its ceiling. An agent embedded inside a single DSP can only buy what that DSP sees, on the terms that DSP transacts, with the signals that DSP ingests. It automates a lane. What buyers actually need is orchestration across lanes: a buying architecture that transacts agent-to-agent where agent-to-agent is available, sources programmatic demand where the bidstream is the only road in, and allocates budget between them based on which is producing the declared outcome. That requires a platform that is AI-native at its core and protocol-agnostic at its edges.
How ORCA delivers value across both paths
ORCA (Orchestrated, Real-time, Collaborative Agents) is Adgentek's agentic media buying platform, built natively on AdCP, and designed from the start to buy across both architectures rather than inside one. A strategist declares the business outcome, such as CPA, ROAS, or cost per qualified lead, along with budget and guardrails. ORCA's coordinated agents handle the rest: planning agents propose allocation, buying agents transact, optimization agents reallocate in real time, and measurement agents close the loop against the outcome.
On the AI-native path, ORCA is a first-class citizen. Its buying agents are dispatched directly to AdCP-compliant sell-side infrastructure, including Adgentek's own Agentic Ad Server, negotiating and executing deals agent-to-agent with no bidstream in the middle. That is the compressed chain the holding companies are chasing: fewer intermediary layers, less information loss, and direct access to inventory that never appears in an auction, from AI surfaces to agent-represented CTV and publisher direct.
On the programmatic path, ORCA does not pretend the existing ecosystem is going away. Where the outcome lives in RTB-cleared inventory, ORCA's agents source it through OpenRTB pipes, applying the same outcome logic and the same governance to programmatic buys as to agent-to-agent ones. The point is side-neutral, protocol-agnostic orchestration: budget follows the outcome, not the plumbing. And because ORCA also receives expressed-intent and engagement signals from AI surfaces, including interaction depth from formats like Spark, its optimization runs on intent data that no bidstream-only platform has, and applies it across every channel it buys.
The third piece is the one the past month proved matters most: judgment. WPP shipped its buyer agent with mandatory human approval, and BCG's finding that only 8 percent of CMOs run truly autonomous campaigns is a trust gap, not a technology gap. ORCA's answer is architectural. Every agent action is a structured proposal that passes through a judgment layer of specialist judges, covering authorization, evidence, exposure, policy, and reversibility, before anything executes. Each proposal resolves to one of four outcomes: allow, block, revise, or escalate to a human. Decisions are logged with memory provenance, so every buy is auditable back to the signal that justified it. That governance model works identically whether the underlying transaction is an AdCP deal or an OpenRTB buy, which is exactly what makes autonomy trustworthy enough to scale across both paths.
One thing ORCA is not: a DSP. It does not run auctions, resell inventory, or ask strategists to manage bids and line items. It orchestrates outcome-driven agents that transact wherever the outcome is best served, and it treats DSP-cleared inventory as one road among several rather than the destination.
How to think about what comes next
The two paths will run in parallel for years, and the protocol layer is where they meet: AdCP is explicitly designed to coexist with OpenRTB, and hand-offs between the architectures will become routine. But the direction of travel is one way. Every month adds AdCP-compliant sell-side endpoints, agent-negotiated deal volume, and AI surfaces generating expressed intent, while nothing adds new reasons to route a buy through five intermediaries. Buyers who adopt path one tooling alone will get more efficient at the old model. Buyers who adopt an AI-native orchestration layer get the efficiency and a structural position in the new one. For agencies and brands, that is the real decision on the table this year, and it is a much better question than which platform's copilot to turn on.
Getting started
If you are a brand or agency evaluating agentic media buying, the practical first step is a pilot that spans both paths: outcome-based buying into AI surfaces alongside your existing programmatic channels, with ORCA orchestrating and judging every action. Reach out at hello@adgentek.ai and we will scope it against your current media mix.
