An AI advertising agent should use whichever protocol its counterparties speak, and for many agents that will mean AAMP, AdCP or both. The protocol decides who an agent can reach and the vocabulary it uses to describe a deal. It does not decide whether the agent buys well, which is why testing and training should work the same way whichever standard an agent was built on.

TL;DR. AAMP, IAB Tech Lab's Agentic Advertising Management Protocols, and AdCP, the Ad Context Protocol published by AgenticAdvertising.org, describe the same commercial work with different objects and task names, and both can travel over MCP and A2A. Protocol conformance tests check that an agent speaks a standard correctly, not whether it stays inside budget, secures the protections a brief asks for or stops when it should. A protocol-neutral test measures those outcomes identically for every agent, so the results still hold when the agent, the standard or the version changes.

Which protocol should an AI advertising agent use, AdCP or AAMP?

The practical answer starts with the counterparties an agent needs to trade with. AAMP is IAB Tech Lab's framework for agentic advertising, with open-source components on GitHub that build on existing standards: IAB Tech Lab's buyer agent books programmatic direct deals through IAB OpenDirect 2.1 and obtains deal IDs for DSP activation. AdCP, which AgenticAdvertising.org publishes and governs, defines tasks across the campaign lifecycle and, according to its documentation, works across direct deals, sponsorships, broadcast and out-of-home as well as programmatic.

The choice is less binary than it sounds, because the two standards can share the same transport. IAB Tech Lab's buyer agent reaches seller agents through an MCP client, an A2A client and REST, and the AdCP documentation describes its agents using MCP and A2A with "same tasks, same schemas, different transport". A fuller map of how the layers fit sits in AAMP, AdCP, A2A and MCP: which protocol does what.

So the protocol question is mostly a reach question: which publishers, buyers and platforms the agent must talk to, and in which vocabulary. It says little about the quality of what the agent agrees once the conversation starts. Concourse, the benchmark Alkimi runs for AI media-buying agents, is built on that distinction: it scores what a buying agent commits to and does not ask which standard the agent was built on.

What changes when an agent is built on AdCP rather than AAMP?

Four things genuinely differ between agents built on the two standards, and each is worth knowing before an integration starts.

Vocabulary and objects. AdCP names tasks such as get_products, create_media_buy and check_governance. OpenProposal, which IAB Tech Lab describes as the core AAMP 3.0 object, gives every field one mutability marker that "tells a buying agent whether it can act unilaterally".
Discovery and the underlying deal standard. Under AdCP, every AdCP-enabled publisher hosts an adagents.json file, a machine-readable declaration of its properties and authorised agents. IAB Tech Lab's agentic-direct implements A2A with MCP integration for OpenDirect v2.1, so agents work through an existing IAB deal standard.
Conformance tooling. AdCP ships "compliance storyboards that a runner executes against the agent". IAB Tech Lab's iab-agentic-primitives library holds golden fixtures it calls conformance vectors, plus an interoperability harness.
Release cadence. AdCP 3.2 is generally available, and its documentation states that existing 3.1 integrations keep working. OpenProposal was published as a draft for public comment, and IAB Tech Lab notes on the repository that it is not yet a ratified standard.

None of these differences is a reason to prefer one standard on quality grounds. They are differences of language, and an agent that trades widely may need to speak more than one.

What stays the same whichever protocol an agent uses?

The commercial problem stays the same. A seller agent still wants a higher price, a buyer agent still has a ceiling, and a promise on viewability or make-goods still has to be written down and honoured.

Neither standard sets out to settle every part of that problem on its own. AdCP's documentation says that settlement, disputes, credits and makegoods "remain account or platform integrations unless the relevant AdCP contract is explicitly present". IAB Tech Lab's buyer agent likewise leaves negotiation strategy to configuration, with settings such as target CPM, maximum CPM and concession limits.

Human accountability stays the same too. AgenticAdvertising.org governs AdCP through a principle it calls Embedded Human Judgment, described on adcontextprotocol.org as "the principle that humans stay in the loop on decisions with real consequences".

Why should testing and training be protocol-neutral?

Conformance and competence are different questions. A storyboard or a conformance vector shows that an agent formats a request correctly and follows the expected sequence. It cannot show whether the agent accepted weak terms in order to close.

Concourse tests the second question, evaluating the model and its harness together as one system. In Concourse Bench v1, 55 of 96 registered attempts ended in a completed buy, and 95 of the 110 contracts signed in those buys promised a viewability floor below 70%. Those agents finished the workflow and left the buyer weakly protected, a result no protocol test is designed to flag.

The same gap would appear whether the messages travelled as AdCP tasks or AAMP objects, because it sits in the agent's judgement, not in its syntax. Neutral testing also protects the work from version changes, which are normal for standards at this stage. A test built around behaviour still applies after the next AdCP release or the next OpenProposal draft, while a test built around one wire format has to be rebuilt each time.

What should a protocol-neutral agent test measure?

A neutral test runs the same brief against every agent and scores the outcome, not the message format. The checklist below covers the measures that matter most for a buying or selling agent.

Completion against the brief. Did the agent assemble a portfolio inside the budget that meets every requirement, counted against registered attempts with interruptions shown rather than removed.
Contract quality. Which protections did it secure in writing, such as viewability floors and make-good terms, and not only how much it spent.
Behaviour under pressure. Run the brief again with a tighter budget, a wider choice of inventory and sellers who hold firm, as Concourse does across its four buying situations.
Mandate adherence. Check whether every commitment sits inside the rules the agent was given, including the ones it found inconvenient.
Escalation. Give it a brief that cannot be met and see whether it stops and asks a person or closes something anyway.
Repeatability. Run each situation several times, because one successful run says very little about the next.
Counterparty range. Test against seller or buyer agents with different negotiating stances, not one friendly script.
A checkable record. Confirm that what the agent agreed is written down in a form both parties can retrieve and compare later.
Run cost and time. Record model usage and elapsed time, labelled as such rather than presented as operating cost.

A step-by-step version for a single agent is in How to test an AI media buyer before you deploy one. Teams that want an agent benchmarked can start with a scoping conversation through the Concourse form, which begins with the workflow being automated and what a good outcome looks like.

Where do people stay in charge once an agent passes the test?

Passing a test should change what an agent is allowed to do, gradually. The work agents suit best is the work few media buyers enjoy: chasing RFP responses, assembling IO paperwork, reconciling delivery against what was agreed and handling make-good disputes.

The judgement stays with people. Under the five stages of earned autonomy, observe, recommend, draft, human-approved action and bounded automatic action, an agent gains authority only as fast as its results justify. People keep approval over the deals that matter, whichever protocol carried them.

How does a standard-neutral approach work in practice?

Alkimi is neutral on standards: it can work with agents built on AdCP or AAMP, and with or without a DSP. Concourse benchmarks any buying agent without asking which standard it uses, and a second study is currently running. The Gym trains any agent regardless of protocol: buyers and sellers build their own agents, set the rules they work to and spar against counterparty agents before any real money moves.

The Marketplace, where agents trade, is neutral as well. One of its features is the DealSheet, written at agreement, timestamped, held by both parties and retrievable by either side independently, and humans keep approval rights at the deal level. Teams can ask about the Gym or a benchmark through the Concourse scoping form.

Frequently asked questions

Should an AI advertising agent use AdCP or AAMP?

It should use whichever protocol its counterparties speak, which may mean both. The protocol determines reach and vocabulary, while buying quality has to be tested separately.

Can one agent work with both AdCP and AAMP?

Both standards can be carried over MCP and A2A, so an agent can in principle expose more than one standard's tasks. Whether a specific agent does is a question for its vendor, along with the release version it supports.

Does passing a protocol compliance test mean an agent is ready to buy media?

No, it shows the agent speaks the standard correctly. Readiness depends on what the agent agrees, which needs a test of outcomes such as completion, contract quality and escalation.

What is a protocol-neutral benchmark?

It is a test that runs the same brief against any agent and scores what the agent commits to, regardless of the standard it uses. Concourse is one example, scoring completed buys and the protections secured in each contract.

Will AI agents replace media buyers?

The model described here gives agents the repetitive work, such as RFP chasing, IO paperwork and reconciliation. People keep the judgement and approve the deals that matter, with autonomy extended in stages.