Agentic AI is software that receives a goal rather than an instruction, breaks that goal into steps, uses external tools to carry out those steps, and keeps working without a human approving each action. In advertising the goal might be a campaign brief, and the steps might include pulling inventory availability, checking audience segments, negotiating a price, creating a deal and reporting on delivery. The phrase exists because large language models became capable of planning and calling software, which turned a text generator into something that could operate a media buying stack.

The term is used loosely, and the looseness has consequences. The OECD drew the sharpest line available in a 34-page working paper published on 13 February 2026. An AI agent, in that framework, is a single system that perceives and acts on its environment with some autonomy, using tools to reach specific goals. Agentic AI refers to systems composed of multiple coordinated agents that decompose tasks, collaborate, and pursue complex objectives over extended periods with minimal supervision. One is a component. The other is an architecture. The paper found the two terms used interchangeably across industry, academia and policy, which matters because liability, oversight and technical specification all depend on which one is being described.

Autonomy is not agency

The OECD paper separates two properties that vendor marketing tends to merge. Autonomy is the capacity to act without direct human involvement. Agency is the capacity to formulate goals, reason over long horizons and adapt strategy. A thermostat has autonomy and no agency.

Four levels of action autonomy follow. At no-action autonomy the system recommends and a human decides. At low-action autonomy it proposes and proceeds only on approval. At medium-action autonomy it acts unless a human intervenes. At high-action autonomy it acts alone. Nearly every advertising deployment shipped so far sits in the first two bands, which is the single most useful fact for reading vendor announcements.

Britain's Digital Regulation Cooperation Forum reached a similar structure by a different route. Its foresight paper, published on 31 March 2026 by four regulators covering competition, financial services, data protection and online safety, sets out five levels of autonomy from reactive tool upward. Retrieving information does not by itself make a system an agent, according to the paper; what counts is whether it can take in data, decide, and act.

The plumbing

Nothing in agentic advertising works without a way for a model to call software it does not own. That layer is the Model Context Protocol, published by Anthropic in November 2024 and later donated to a Linux Foundation fund, which standardizes how a language model discovers and invokes external tools. Its complement, Google's Agent2Agentprotocol, standardizes how one agent talks to another. Advertising adopted MCP first and heavily. Amazon Ads opened a closed beta of its MCP server on 13 November 2025, Adform published 29 read-only skills against its FLOW demand-side platform, and X exposed 23 tools for its ads API, ten of which write to live accounts, with every agent-created campaign starting paused.

Above the plumbing sits the transaction layer, and here the industry has produced two competing answers. The Ad Context Protocol, launched on 15 October 2025 by Scope3, Yahoo, PubMatic, Swivel, Triton and Optable under an MIT license, defines nine core tasks covering product discovery, campaign creation, creative management and optimization, exposed through both MCP and A2A.

IAB Tech Lab's answer is broader. Named AAMP, for Agentic Advertising Management Protocols, on 26 February 2026, it stacks three pillars. At the base, the Agentic Real-Time Framework defines how agent code runs inside auction systems: gRPC with protobuf serialization, Open Container Initiative compliant containers, and a security model prohibiting all network traffic except communication with the orchestrating entity. Each container declares its intents and any auction changes it wants, and the orchestrator can reject them. IAB Tech Lab states the framework cuts latency by 90%, a prerequisite rather than a benefit given that auction decisions must complete inside 100 milliseconds.

The middle pillar carries seven protocol components mapped onto existing standards, among them Agentic Audiences on the User Context Protocol donated by LiveRamp, Agentic Ad Objects on AdCOM, Agentic Direct on OpenDirect, Agentic Deals on the Deals API and Agentic Bid on OpenRTB. Buyer and seller software development kits share a three-tier hierarchy. Level 1 is a Portfolio Manager handling budget allocation on the buy side and yield strategy on the sell side. Level 2 holds channel specialists for connected television, mobile app, display, video and native. Level 3 holds functional agents: research, execution and reporting for buyers; pricing, availability, proposal review and audience validation for sellers. A taxonomy guardrail band constrains all three tiers to shared vocabularies rather than proprietary ones.

The top pillar is an Agent Registry, which reached ten entries on 11 March 2026, all of them MCP servers validated against privacy framework identifiers before approval.

Origin

Intelligent agents are old: Russell and Norvig defined them as systems that perceive and act in an environment decades before the current wave. What changed was inference. Gartner named agentic AI its top strategic technology trend for 2025 on 21 October 2024, predicting that by 2028 at least 15% of day-to-day work decisions would be made autonomously, up from zero.

Advertising moved a year later, and fast. Four weeks in late 2025 produced AdCP, LiveRamp's donation of the User Context Protocol, Amazon's Ads Agent and the Agentic RTB Framework. PubMatic launched AgenticOS with live campaigns on 5 January 2026, and Magnite built a seller agent into SpringServe the next day. AAMP 2.0 arrived in April 2026 and version 2.3 on 30 July.

Why it matters

Two-thirds of advertisers were prioritizing agentic AI for campaign execution by January 2026, according to IAB research published that month. The stated case is compression. Yahoo DSP's always-on delivery agent diagnoses underpacing across targeting, supply and settings, then recommends or, with approval, executes fixes.

Money has not followed at the same rate. Magnite chief executive Michael Barrett capped 2027 agentic spend forecasts between negligible and $700 million in July 2026 and described the market in August as one buyer transacting with one seller rather than one-to-many. Holding companies have meanwhile begun reselling model capacity to clients at a markupwhile transacted agentic media stays thin. PubMatic chief executive Rajeev Goel forecasts 25% of digital advertising executing autonomously by 2028 and 50% by 2030, a vendor projection without independent corroboration.

Where it fails

The independent pricing evidence runs against the efficiency story. DataBeat's June 2026 report found conventional programmatic buyers holding a 13.4% CPM advantage over AI agents, with agentic demand appearing in 86% fewer auctions and clearing $6.13 against $6.95: a television buying pattern imported into real-time infrastructure.

Cost is the second constraint. Agents consume roughly five times the tokens a human does per session, and KPMG research published in August 2026 found 49% of leaders had scaled back rollouts when running costs outran value. One Dutch operator collapsed twelve MCP calls into a single buyer agent to contain that burn.

Gartner predicted on 25 June 2025 that more than 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, and estimated that only about 130 of the thousands of vendors claiming agentic capability were building it. It calls the rest agent washing.

Trust runs through every objection. Augustine Fou, a fraud researcher, told The Current that more automation means less transparency and that agents can still act for people with bad incentives. IAB Tech Lab chief executive Anthony Katsur warned in December 2025 that the discourse had reached a fevered pitch while transparency, privacy and measurement stayed unresolved. UC Berkeley's Center for Long-Term Cybersecurity named agent identity and accountability as foundational gaps in February 2026, and Spain's data protection authority mapped GDPR exposure from agentic workflows the same month, coining BYOAgentic for employees running no-code agents outside governance.

What it is not

AI agent describes a single autonomous system. Agentic AI describes several working together. Vendors use the terms interchangeably; the OECD does not.

Automated bidding predicts the value of one impression and prices it, with no goal decomposition, no tool use and no persistence across a campaign. Smart Bidding is a model inside an auction, not an agent.

Generative AI produces output on prompt. The absence of self-initiated action is the whole distinction.

AI crawlers such as GPTBot and ClaudeBot retrieve pages, and retrieval alone does not make an agent.

Agentic commerce is the consumer-facing case: an assistant buying on a shopper's behalf through Google's Universal Commerce Protocol or OpenAI's Agentic Commerce Protocol. Same technology, opposite side of the transaction.

Recent developments

Consolidation has not happened. IAB Tech Lab operating chief Shailley Singh counted thirteen functions where AAMP and AdCP overlap on 20 August 2026, covering almost the entire arc of a media buy. AdCP still counts neither Google, Amazon, The Trade Desk nor Microsoft among its adopters.

Guardrails are the current product cycle. PubMatic shipped a five-step governance layer restricting what agents can transact inside AgenticOS on 5 August 2026. Fluency disclosed the same month that it blocks agents from executing against live budgets across roughly $3 billion in managed spend. StackAdapt research on 19 August found 6% of marketers act on in-platform AI recommendations. InMobi placed a conversational agent inside its Buyer Hub on 13 August, and Sabio embedded Gentoro agents in DSP campaign workflows at the end of it. Two years after the term entered the vocabulary, most deployments still stop short of letting the machine spend money unattended.

Timeline

  • 21 October 2024: Gartner names agentic AI its top strategic technology trend for 2025
  • November 2024: Anthropic publishes the Model Context Protocol
  • 25 June 2025: Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027
  • 1 October 2025: LiveRamp introduces agentic orchestration for its marketing tools
  • 15 October 2025: Ad Context Protocol launches with six founding members
  • 3 November 2025: LiveRamp donates the User Context Protocol to IAB Tech Lab
  • 11 November 2025: Amazon launches Ads Agent at unBoxed
  • 13 November 2025: IAB Tech Lab opens the Agentic RTB Framework for public comment; Amazon opens its MCP server beta
  • December 2025: Anthropic donates MCP to a Linux Foundation fund
  • 5 January 2026: PubMatic launches AgenticOS
  • 6 January 2026: Magnite announces a seller agent in SpringServe
  • 13 February 2026: OECD publishes "The Agentic AI Landscape and Its Conceptual Foundations"
  • 26 February 2026: IAB Tech Lab names AAMP and publishes its three-pillar architecture
  • 11 March 2026: AAMP Agent Registry reaches ten entries
  • 25 March 2026: Kochava opens an AAMP sandbox in public beta
  • 31 March 2026: UK Digital Regulation Cooperation Forum publishes its agentic AI foresight paper
  • April 2026: AAMP 2.0 adds five transaction types
  • 11 June 2026: Mediaocean, Magnite, Teads and Walmart Connect announce agentic products on the same day
  • 22 June 2026: DataBeat reports a 13.4% CPM advantage for conventional buyers over AI agents
  • 30 July 2026: AAMP 2.3 adds pricing provenance and a vendor approval gate
  • 5 August 2026: PubMatic launches agent guardrails on AgenticOS
  • August 2026: Agentic Real Time Framework 1.0 is finalized
  • 20 August 2026: IAB Tech Lab counts thirteen functional overlaps between AAMP and AdCP

Summary

Who: Standards bodies including IAB Tech Lab and Prebid.org; protocol founders Scope3, Yahoo, PubMatic, Swivel, Triton and Optable; platform operators such as Amazon, Magnite, PubMatic, Yahoo, Adform, InMobi and DoubleVerify; model providers Anthropic, Google and OpenAI supplying the underlying protocols; and regulators including the OECD, the UK's DRCF and Spain's AEPD.

What: Systems of coordinated AI agents that decompose a goal, use external tools and act over extended periods with limited supervision. In advertising this means buyer and seller agents transacting through MCP, AdCP and AAMP against existing standards including OpenRTB, AdCOM, OpenDirect and the Deals API.

When: The term entered mainstream use in October 2024. Advertising infrastructure arrived from October 2025 onward, with AAMP named in February 2026 and shipping version 2.3 on 30 July 2026.

Where: Across programmatic display, connected television, mobile app, audio and digital out-of-home, with IAB Tech Lab extending the work toward linear television and radio.

Why: Buying and selling advertising involves large volumes of repetitive coordination that agents can absorb. Independent evidence for the resulting efficiency remains limited: conventional buyers cleared 13.4% higher than agentic demand in June 2026 measurement, Gartner expects more than 40% of agentic projects to be canceled by end-2027, and most production deployments still require a human before money moves.