PubMatic’s AgenticOS debuted on January 5th as one of the first operating systems for agentic advertising. Launched in partnership with Butler/Till, WPP Media and others, we’ve since seen the most rapid adoption of any product launch in PubMatic history: more than 250 agentic deals transacted, 100+ new entrants to our AI accelerator program, and we just released our first case study with Butler/Till, with more to come.
But what really is an operating system for agentic advertising?
An operating system isn’t a single feature — it’s a coordination mechanism that makes everything else run. PubMatic’s AgenticOS is a system of interconnected agents working across PubMatic’s platform — its premium publisher inventory, 250+ data partners, and advanced infrastructure — to handle every phase of campaign execution, from audience discovery through live optimization, activated through open protocols that connect to any buyer’s preferred AI environment.
The industry has generated a lot of conversation — and a fair amount of confusion — about those protocols. Here’s the short version.
The Protocol Explainer
- AdCP (Ad Context Protocol), co-founded by PubMatic alongside Yahoo, Scope3, and others, defines the advertising-specific vocabulary: how agents use a shared language to discover inventory, initiate media buys, sync creative, retrieve delivery data, and activate audience signals.
- MCP (Model Context Protocol), developed by Anthropic, is the universal adapter that lets any AI system — Claude, ChatGPT, or a proprietary agent — connect to external platforms and call tools natively.
- A2A (Agent-to-Agent), developed by Google, does the same at a different transport layer, better suited for multi-step autonomous workflows and real-time streaming.
- AAMP (Agentic Advertising Management Protocols) and ARTF (Agentic Real Time Framework), developed by the IAB Tech Lab, define how agentic advertising systems are structured, executed, and made interoperable within existing industry standards.
In short, AdCP defines what advertising agents do. MCP and A2A define how they do it. And AAMP defines how agents operate within advertising infrastructure at scale through ARTF for real-time execution, shared schemas for buyer-seller communication, and an agent registry for identity and trust.
These protocols make AgenticOS and its agents fully callable.
Here’s what’s live inside AgenticOS today:
Audience Discovery
PubMatic’s Audience Discovery agent queries audience segments in natural language — no manual taxonomy navigation required. It surfaces relevant segments from across PubMatic’s 250+ data partner ecosystem before any campaign parameters are set, putting the right data in context for every downstream decision.
Inventory Discovery Marketplace
The Inventory Discovery agent takes campaign intent — format, audience, brand safety, targeting criteria — and curates real-time inventory packages from PubMatic’s publisher network. It queries supply, applies filters, and returns a ranked set of options for buyer review. The Butler/Till campaign — which produced a 82% reduction in buy-side costs and 40% lift in impressions — used this agent as part of its CTV targeting workflow.
Campaign Insights
PubMatic’s Campaign Insights agent introduces AI-powered diagnostics to shift optimization from reactive to proactive. Pacing anomalies, delivery gaps, targeting inefficiencies — these surface as they emerge, not after the fact. This agent works with buyer agents to query live campaign performance data directly from within its native LLM environment.
Media Activation
This is where AgenticOS moves decisively beyond what the protocol standards alone define. The Media Activation agent extends the full capabilities of PubMatic’s media bidder, Activate, to be agentic. This agent includes budget strategy, pacing type, campaign flighting, campaign goals, conversion events, brand safety, viewability standards, content targeting, day parting, and deal targeting across PG, PMP, and open exchange, and specific publisher selection across PubMatic’s 2,000+ premium publisher network.
An agent can do everything a media trader does — without leaving its native AI environment. Once the Media Activation agent takes over campaign execution, traders can shift their focus to client service and strategy while the agent manages the day-to-day workflow.
Buyer Agent
Our Buyer Agent acts as the advertiser’s central coordinator within AgenticOS. It takes high-level, natural-language goals — like “Run a CTV campaign targeting sports fans” — and translates them into clear actions. It then activates the appropriate agents, such as Audience Discovery, Inventory Discovery, Campaign Insights, and Media Activation, guiding each step of the workflow and bringing the results back into the buyer’s preferred LLM.
What’s Powering It
AgenticOS runs on NVIDIA GPU-powered infrastructure purpose-built by PubMatic for advertising inferencing. Campaigns activate directly at the supply source — no intermediary layer, no QPS constraints. What the agent requests, the bidder executes.
This is what it means to build AI into the infrastructure rather than layer it on top. The shift from machine learning to generative AI isn’t just a software change — it requires a fundamentally different compute foundation, one capable of processing the intent-driven, real-time decisions that agentic workflows demand.
PubMatic’s fully owned and operated GPU environment — built through a multi-year co-innovation partnership with NVIDIA — means every inference, every signal, every autonomous decision an agent makes runs on dedicated, trusted computing. No shared cloud infrastructure. No third-party dependencies at the critical layer. When the industry moves to agentic at scale, the platforms with purpose-built infrastructure are the ones buyers and publishers can rely on.
The Bigger Picture
Connectivity and interoperability is just the entry point. What determines whether agentic buying actually works isn’t just the depth of what an agent can do once it’s connected — it’s whether buyers can trust it to do those things accurately, reliably, and consistently. Functional depth matters. So does the integrity of every inference, every signal, every autonomous decision the system makes.
That thinking has been foundational to how we’ve built AgenticOS from the start. Every exchange is structured and logged, enabling auditability and strengthening partner trust. Agents must authenticate each exchange, access only approved datasets, and leave transparent audit trails.
This isn’t a governance layer bolted on after the fact — it’s been core to PubMatic’s agentic architecture since we published our first open agent-to-agent specification in 2025.
Trust in agentic systems isn’t declared. It’s earned — through consistent, auditable performance over time. AgenticOS was built with that in mind.
The infrastructure is open. The capability is complete. The dollars are flowing.
Connect with your PubMatic representative to explore what AgenticOS can do inside your current AI workflow.
