Agentic advertising is the use of AI agents that make and act on media-buying decisions autonomously, within human-set boundaries, rather than executing decisions a human has already made. The vocabulary around it is new, contested, and being written in real time by standards bodies, platforms, and holding companies at once. This glossary defines the terms a programmatic team will meet, in plain language, with the distinctions that actually matter for evaluating what a vendor is claiming.
TL;DR. The terms below cluster into four groups: what an agent is and does (agent, agentic advertising, agentic marketplace), how agents talk and transact (agent-to-agent negotiation, MCP, AAMP, AdCP), how autonomy is bounded (earned autonomy, human-in-the-loop, guardrails), and what has to be true for any of it to be trustworthy (shared record, auditability, supply path optimisation). The single distinction to carry through all of them: automation executes a decision, an agent makes one. Everything contested in the category traces back to that line.
Agentic advertising
The use of AI agents to make autonomous media-buying decisions within boundaries a human sets. The defining feature is not speed or automation, both of which programmatic already delivered, but delegation of the decision itself. Under programmatic, a human decides the strategy and software executes it. Under agentic advertising, the human sets a goal and constraints, and the agent decides how to reach the goal, including choices the human did not explicitly anticipate. The distinction matters because it changes what has to be auditable: not just whether an ad ran inside the rules, but why the agent made the calls it made.
Agent (AI agent)
Software that runs a decision loop on a buyer's or seller's behalf: it reads the current state, decides on an action toward a goal, takes the action, and reads the result. An agent differs from a bidding algorithm in scope. A bid algorithm adjusts one variable inside fixed rules; an agent reasons across many levers at once, targeting, budget, bidding, pacing, inventory, and can act on the relationships between them. In the emerging specifications, agents are often structured in a hierarchy: a strategy-level agent that sets direction, channel-level agents for CTV, display, video and the like, and functional agents that perform bounded tasks such as pricing or audience planning.
Agentic advertising marketplace
A venue where buyer agents and seller agents discover each other, negotiate terms, and transact directly, rather than where humans configure campaigns that software executes. It differs from a demand-side platform, which is buying software a human operates, and from an ad exchange, which is an auction that clears single impressions. The marketplace's distinct job is to govern how autonomous parties negotiate and to prove what they agreed, because agentic deals persist and evolve across a campaign rather than clearing in a single moment.
Agent-to-agent (A2A) negotiation
The process by which a buyer's agent and a seller's agent agree the terms of a deal directly, without a human negotiating each point. This is the capability that separates an agentic marketplace from an exchange: not bidding into an auction, but negotiating inventory, price, volume, and conditions. The first cross-platform agent-to-agent media buys have already run in production, with buy-side and sell-side agents negotiating over a shared communication protocol. That said, the people writing the standards are explicit that agentic buying is not yet running at scale, and A2A negotiation is also where governance concentrates, because an agent agreeing terms is committing a party to something, and the boundaries on what it may agree to are the difference between a controlled deal and an uncontrolled one.
Model Context Protocol (MCP)
An open protocol, drawn from the wider AI field rather than invented for advertising, that standardises how an AI model or agent connects to external tools, data, and other agents. In agentic advertising it is one of the communication layers the standards bodies are building on, so that a buyer's agent and a seller's agent can exchange information in a common language rather than through bespoke one-to-one integrations. When the industry's early agent registries first expanded, the agents listed were MCP-based, which is why the term appears so often in agentic announcements.
AAMP (Agentic Advertising Management Protocols)
The IAB Tech Lab's umbrella initiative for agentic advertising standards. Rather than build a new framework from scratch, AAMP extends established advertising standards, OpenRTB, AdCOM, OpenDirect, VAST, and the Deals API, into the agentic layer, and adds agent communication protocols on top. The argument for extending rather than replacing is that existing standards encode knowledge refined through billions of transactions, so building on them is faster and less fragmenting. AAMP includes open-source buyer and seller agent reference implementations and, in its current versions, guardrails that require human approval on any path that commits spend above a value threshold.
AdCP (Ad Context Protocol)
A separate open initiative defining its own approach to agent-to-agent communication and transactions in advertising, backed by a coalition including Scope3, PubMatic, and Yahoo and governed by the non-profit AgenticAdvertising.org. AdCP and AAMP are the two efforts competing to become the common technical layer for agentic advertising, which is why the pairing appears in category coverage. For a buyer, the contest matters less than its outcome: if one becomes the shared standard, agentic buying scales across the ecosystem; if the market splinters between them, the interoperability that makes agentic buying efficient is harder to reach.
Earned autonomy
A model in which an agent's freedom to act without human approval expands gradually as its record of good decisions accumulates, rather than being granted all at once. In practice this looks like a graded scale: the agent begins in recommend-only mode, where a human approves each action, and the band of actions it can take unsupervised widens as trust is established, with low-stakes reversible actions automated first and high-stakes irreversible ones held behind approval longest. Earned autonomy is the counter to the assumption that agentic means fully autonomous. The durable deployments treat autonomy as something an agent earns against an audit trail, not a switch that is flipped on.
Human-in-the-loop
An arrangement in which a human must approve certain agent actions before they take effect. In agentic advertising the loop is usually placed by stakes: actions that commit spend, open new deal terms, or are hard to reverse require sign-off, while minor reversible adjustments do not. The current industry specifications build this in by design, requiring human approval on spend-committing paths above a set value. The human-in-the-loop is not a courtesy. It is the interrupt that prevents the failure mode specific to autonomous systems, in which a single misread signal produces the same bad action repeatedly at speed.
Guardrails
The predefined limits within which an agent is allowed to operate: spend caps, approval thresholds, brand-safety requirements, and rules constraining what an agent may negotiate or agree. Well-built agentic systems ask the buyer to declare goals, guardrails, and brand-safety conditions up front, and then let the agents act only inside those declared boundaries. Guardrails are deterministic where the agent's reasoning is not: they are the fixed limits that make an unpredictable decision-maker safe to deploy. An ungoverned agent can optimise toward a flaw at frightening speed, eroding yield to hit a fill target or misapplying pricing, which is why the guardrails, not the agent's cleverness, are what a buyer should scrutinise first.
Shared record (shared deal object)
A single, verifiable record of a deal that both the buyer's agent and the seller's agent reference, as opposed to each keeping its own separate record. This is the term the category is quietly organising itself around, because it names the fix for agentic advertising's central risk. When two agents keep separate books of the same deal, those books drift apart over a campaign, and every later decision is built on a divergent version of events. A shared record removes the drift by giving both sides one source of truth to read from and write to. The people writing both competing standards agree on the underlying point: a model asserting that a transaction occurred is not the same as a deterministic, auditable record of it. The concept has direct precedent in financial markets, where a shared settlement layer replaced two parties reconciling separate trade records forever after the fact.
Auditability
The property of a system in which any decision can later be traced to the reasoning that produced it, in a form a human can inspect. In agentic advertising, auditability is harder and more important than in programmatic. Harder because an agent's decision draws on many signals and a chain of reasoning rather than a single rule that either fired or did not. More important because the buyer has to be able to explain, months later, why an agent moved budget or agreed terms, to someone who will not accept "the agent handled it." A system that reports outputs and a summary, with no recoverable logic in between, is not auditable in the sense that matters. This is precisely the gap the standards bodies describe when they say natural-language confirmation is not a transaction record.
Supply path optimisation (SPO)
The practice of choosing the most direct, transparent, and efficient route for a bid to reach inventory, minimising unnecessary intermediaries and hidden fees. SPO predates agentic advertising, but agents change it: rather than accepting whatever path the pipeline serves, an agent can actively favour paths with better fee visibility and quality signals as part of its optimisation. SPO connects to the category's wider transparency problem, since the same opacity that lets fees hide in a supply chain also makes an agent's decisions harder to verify. An agent choosing not just what to buy but how to buy it is SPO moved from a periodic human exercise to a continuous automated one.
Made for advertising (MFA)
Low-quality websites and apps built primarily to carry ads rather than to serve an audience, long a source of wasted programmatic spend. MFA is included here because it is the benchmark problem agentic curation is often claimed to address: an agent selecting inventory on quality signals should, in principle, route spend away from MFA more consistently than periodic human review does. The claim is plausible but not automatic, and it depends entirely on the quality signals the agent can read and trust, which returns the question to the record and the data underneath the decision.
Additive versus extractive
A test for any new layer in the advertising supply chain: does it add value that reaches buyers and publishers, or does it extract value on the way through? The framing is useful for agentic advertising specifically because an agent is neither by nature. An agent acting on data it can verify against a shared record is additive: it makes faster, defensible decisions and leaves a record a human can check. An agent acting on data it cannot verify is extractive: it makes fast decisions that look clean until someone needs to prove what happened. The distinction is not about the agent's intent, which it does not have, but about the infrastructure it runs on.
How to use this glossary
The terms are not neutral. Each one is being defined in public by parties with an interest in where the definition lands, which is why a programmatic team benefits from holding the plain meaning rather than the marketed one. When a vendor says its platform is agentic, the useful follow-up is which of these it actually provides: does the agent negotiate or only bid, is its autonomy earned or assumed, are the guardrails deterministic, and is the deal recorded somewhere both sides trust. The vocabulary is the entry point. The distinctions inside it are how a buyer tells a real agentic system from a relabelled one, and, as the reporting on what agentic buying can actually prove today makes clear, the gap between the two is currently wide.
This glossary references the IAB Tech Lab's published agentic advertising standards and the Ad Context Protocol initiative for the technical terms, and industry analysis of agentic governance for the autonomy and guardrail definitions. The shared-record entry draws on published simulation research into agentic deal reconciliation conducted by Alkimi; that research models the current specification architecture and is not an assessment of any specific production platform.