The most dependable way to learn about agentic advertising standards and research is to read the primary sources: IAB Tech Lab's AAMP pages and GitHub repositories, the AdCP documentation from AgenticAdvertising.org, the open A2A and MCP specifications, and a small shelf of research papers on agents that plan, negotiate and trade. This list collects 25 of them, each with a short note on what it contains and who should read it.

TL;DR: Start with AAMP and its GitHub hub, because that is where IAB Tech Lab publishes its agentic specifications and reference agents. Read AdCP next, then the A2A and MCP specifications that both frameworks draw on. Finish with the research papers, which test whether language model agents can complete tasks and negotiate under rules, and the risk frameworks that cover prompt injection and disclosure.

How is this reading list organised?

The 25 sources sit in five groups: IAB Tech Lab's AAMP framework, AdCP, the general agent protocols, academic and lab research, and risk and disclosure guidance. Every entry is a primary source, meaning a specification, a code repository, a paper or an official announcement, rather than commentary about one. Where a term is unfamiliar, the Agentic Advertising Glossary covers the vocabulary.

Primary sources matter here because the specifications are still moving, and summaries date quickly. One worked example of using them directly: Alkimi, an enterprise platform for agentic advertising, read the OpenProposal draft in its repository and filed comments and change requests through the public issue tracker. That route is open to anyone with a GitHub account, and reading the draft itself is the only way to know what it actually says.

Which IAB Tech Lab sources define AAMP?

AAMP is IAB Tech Lab's Agentic Advertising Management Protocols, the umbrella under which its agentic work is developed. These nine sources cover the framework, its newest specification and the open-source reference code.

1. The AAMP overview page. IAB Tech Lab's AAMP page describes AAMP as an umbrella initiative for bringing agentic AI into advertising "in a responsible and scalable fashion", built on the standards and taxonomies the industry already uses. It also records that AAMP 2.0 added five transaction types through support for the Deals API, including Programmatic Guaranteed, Preferred Deals and PMPs.

2. The AAMP GitHub hub. The AAMP repository is the index to every AAMP repository, with notes on how each is versioned and governed.

3. The AAMP 3.0 announcement. IAB Tech Lab's press release of 22 September 2026 introduced OpenProposal and connected it to AdCOM, OpenDirect and the Deals API. "OpenProposal fills a major gap between a buyer's brief and a buy commitment, giving buyer and seller agents a standardized way to exchange and evaluate opportunities," said Anthony Katsur, CEO of IAB Tech Lab. IAB Tech Lab has since extended the public comment period to 10 November 2026.

4. The OpenProposal repository. The OpenProposal repository holds the draft specification, worked examples and the comment process. Its README states that commenters do not need to be IAB Tech Lab members, and that the draft is not yet a ratified standard. A plain-English summary of the comments so far sits in OpenProposal v3.0: What the Public Comment Period Raised.

5. The Buyer Agent reference implementation. IAB Tech Lab's buyer agent is open-source code for DSPs, agencies and advertisers to automate programmatic direct buying using OpenDirect 2.1. It browses seller media kits, negotiates with configurable strategies and obtains Deal IDs for activation, talking to sellers through MCP and A2A clients.

6. The Seller Agent reference implementation. The seller agent is the supply-side counterpart for publishers and SSPs, built on OpenDirect 2.1, the Deals API v1.0 and sellers.json. Its README lists human-in-the-loop approval gates and an order state machine running from draft to booked, delivering and complete.

7. Agentic Direct. The Agentic Direct repository is a schema-driven implementation of the A2A protocol with MCP integration for OpenDirect v2.1. The AAMP hub describes it as the core communication and orchestration layer for agent interaction.

8. The Agentic Real Time Framework (ARTF). The ARTF repository covers agent services that run inside a host platform, which the orchestrating platform can call directly. Its build tooling generates code from the OpenRTB 2.6 definitions, so it suits engineers who work at bid time.

9. Agentic Audiences. The Agentic Audiences repository, formerly the User Context Protocol, began as LiveRamp's initial proposal and was open-sourced for the community to develop. It sets out how agents exchange embeddings that encode identity, contextual and reinforcement signals in a privacy-preserving format.

Where is AdCP documented?

AdCP is the Ad Context Protocol, published by AgenticAdvertising.org. Two sources cover it.

10. The AdCP home page. AgenticAdvertising.org describes AdCP as an open protocol that lets AI agents discover, plan, buy, sell and measure media across channels, and says the organisation certifies practitioners and governs how the protocol evolves. The home page currently announces AdCP 3.2 as generally available.

11. The AdCP repository and documentation. The AdCP GitHub repository splits the protocol into a negotiation layer (Media Buy, Creative, Signals, Accounts, Governance and more) and a serve-time decisioning layer. Planning-time tasks run over MCP and A2A, and the full task reference lives in the AdCP documentation. For how AdCP and AAMP relate, see AAMP, AdCP, A2A and MCP: Which Protocol Does What.

Which general agent protocols sit underneath?

Both advertising frameworks draw on two general-purpose protocols. Reading the originals explains design choices that the advertising specifications take for granted.

12. The A2A announcement. Google announced the Agent2Agent protocol on 9 April 2025, with support from more than 50 technology partners. The post explains the aim: letting agents built by different vendors communicate, exchange information securely and coordinate actions.

13. The A2A specification. The A2A specification defines the core operations, from sending a message to getting, listing and cancelling tasks. The A2A GitHub repository holds the source.

14. The MCP announcement. Anthropic open-sourced the Model Context Protocol on 25 November 2024 as a standard for connecting AI assistants to the systems where data lives. The post explains the client and server model that the IAB Tech Lab reference agents also use.

15. The MCP specification. The MCP specification sets out the authoritative protocol requirements, including tools, resources, prompts and a section on security and trust.

What does research say about agents that plan, negotiate and trade?

Most of these papers are arXiv preprints, so read the method before quoting the result. All of them test general agents rather than media buying.

16. ReAct. ReAct, submitted in October 2022, showed language models producing reasoning traces and actions in an interleaved way. It is the pattern behind many agents that plan a step, call a tool and update the plan.

17. τ-bench. τ-bench tests agents that must follow domain rules while talking to a simulated user. Its authors report that even state-of-the-art function-calling agents such as gpt-4o succeeded on fewer than 50% of tasks, and introduce a metric, pass^k, to measure consistency across repeated trials.

18. Evaluating Language Model Agency through Negotiations. Davidson and colleagues, in a paper accepted to ICLR 2024, used negotiation games to test six language models in self-play and cross-play. They found cooperative bargaining hardest, and that even the strongest models sometimes lost to weaker opponents.

19. NegotiationArena. NegotiationArena is a framework for testing language model agents in ultimatum games, trading games and price negotiations. The authors found that by pretending to be desolate and desperate, models could improve their payoffs by 20% against a standard GPT-4.

20. Magentic Marketplace. Magentic Marketplace, from Microsoft researchers, simulates a two-sided market where assistant agents represent consumers and service agents represent businesses. The paper reports that all models tested showed severe first-proposal bias, creating 10 to 30 times advantages for response speed over quality, and the Microsoft Research blog post summarises it.

21. Project Vend. In Project Vend, published in June 2025, Anthropic and Andon Labs had Claude Sonnet 3.7 run a small automated shop in Anthropic's office for about a month. It is a readable account of an agent handling real money, stock and customers.

Which sources cover risk, security and disclosure?

An agent that reads a seller's proposal is also reading input it does not control. These four sources cover that risk and the disclosure rules around AI in advertising.

22. Indirect prompt injection. Greshake and colleagues showed that instructions planted in retrieved data can manipulate an application's functions and control which APIs it calls. Read it before letting an agent process documents written by a counterparty.

23. OWASP Top 10 for LLM Applications. The OWASP list for 2025 puts prompt injection first, as LLM01, and also covers sensitive information disclosure, supply chain risk and data poisoning.

24. NIST AI Risk Management Framework. NIST's framework, released on 26 January 2023, is voluntary guidance for building trustworthiness into the design, use and evaluation of AI systems. NIST has also published a companion playbook.

25. IAB AI Transparency and Disclosure Framework V2. IAB's V2 framework sets out a risk-based, materiality-driven approach to when and how AI involvement should be disclosed in consumer-facing advertising.

What should a reader do after the reading list?

Read one specification end to end, then test a claim against the code that implements it. Standards are only as useful as the evidence that agents can work within them, which makes benchmarks built for media buying a natural next step after the general research above.

Alkimi publishes one such test through Concourse: in Concourse Bench v1, 55 of 96 runs completed a buy, and 95 of 110 signed contracts promised a viewability floor below 70%. For how standards move from draft to adoption, Who Governs Agentic Advertising Standards traces the process.

Frequently asked questions

Where should a newcomer to agentic advertising standards start?

Start with IAB Tech Lab's AAMP overview page and the AAMP GitHub hub, which link to its specifications and reference agents. Then read the AdCP home page and the A2A and MCP specifications that both frameworks draw on.

Are AAMP and AdCP the same thing?

No. AAMP is IAB Tech Lab's framework of agentic protocols, while AdCP is the Ad Context Protocol published by AgenticAdvertising.org.

Can anyone comment on OpenProposal?

Yes. The OpenProposal README says commenters need only a GitHub account, not IAB Tech Lab membership, and the comment period now runs to 10 November 2026.

Are the research papers peer reviewed?

Most are arXiv preprints, which are public but not necessarily peer reviewed. The arXiv page for Evaluating Language Model Agency through Negotiations notes its acceptance to ICLR 2024.

Do the reference agents handle human approval?

The IAB Tech Lab seller agent README lists human-in-the-loop approval gates with configurable conditions. Each team still decides where in its own workflow a person must sign off.