IAB Tech Lab this week released version 2.3 of its Agentic Advertising Management Protocols, adding a pricing provenance field designed to stop AI buying agents from fabricating bid prices when real market data is unavailable.

The release, announced in a note to industry outlets from IAB Tech Lab and authored by Shailley Singh, the organization's COO and EVP of Product, follows the framework's second major version in April, which had introduced five new transaction types including agentic private marketplaces and Programmatic Guaranteed deals. According to IAB Tech Lab, the gap between those two releases produced a wave of activity inside the AAMP GitHub repository, where developers across the advertising industry submitted use cases, flagged bugs during real-world testing, and pushed for capabilities that the earlier version lacked.

That activity shaped what AAMP 2.3 became. Rather than adding features for their own sake, the update concentrates on three areas that determine whether an autonomous buying or selling agent can be trusted with a live advertising budget: how it deploys inside enterprise infrastructure, how it proves the prices it negotiates are real, and how it handles audience targeting in a way that supply-side and demand-side platforms can verify.

What changed inside AAMP 2.3

Enterprise deployment and system interoperability

According to IAB Tech Lab, the Buyer and Seller Agents built on AAMP now have a full deployment surface on AmazonBedrock AgentCore Runtime, supporting both REST and MCP operating modes. The organization reports that the agents carry S3-backed inventory adapters, use CloudFormation templates for provisioning, and have passed nearly 5,000 unit tests. That volume of testing matters because it signals the framework has moved past isolated proof-of-concept code and into something organizations can audit before putting it into production.

Interoperability between agents received separate attention. IAB Tech Lab restored Streamable HTTP as the primary transport mechanism for the Model Context Protocol layer connecting agents, while keeping legacy Server-Sent Events support alive for older clients still running on it. The client now negotiates automatically between the two, removing a manual configuration step that previously required advance coordination between counterparties.

The Seller Agent picked up a security upgrade as well. IAB Tech Lab states that bespoke credential paths for Streaming Hub and Buyer Cloud connections have been replaced with OAuth protocols, moving the framework away from custom authentication schemes toward a standards-based approach. Agents can also scale horizontally across multiple instances rather than remaining tied to a single node, following the addition of pluggable storage backends, including SQLite, Redis, and a combined Postgres-plus-Redis configuration.

On the model side, AAMP 2.3 widens the pool of large language model providers an agent can run on. The framework now supports OpenAI, Google Gemini, and other OpenAI-compatible services, including NVIDIA NIM, Ollama, and Hugging Face. IAB Tech Lab also added execution-layer integrations for Google Ad Manager Reporting and Meta Ads, expanding the tools an agent can call once it has decided what action to take.

A fix for fabricated prices

The second theme running through the release addresses what IAB Tech Lab describes as CPM hallucination: a failure mode in which earlier agent versions could generate a fabricated cost-per-thousand-impressions value when actual pricing data was missing from a negotiation. Because a CPM figure that an agent invents looks identical, on the surface, to one grounded in real market conditions, this was flagged as a dangerous problem for any system negotiating with genuine advertising spend.

AAMP 2.3 addresses the issue by adding a field that records pricing provenance throughout a negotiation. With that field present, an organization can trace where each price came from at every step of a deal and set guardrails preventing the agent from inventing floor prices outright. IAB Tech Lab frames the design philosophy directly: an agent capable of stating it does not know a price, rather than confidently generating one, is the kind of system that can be deployed and trusted afterward.

That trust mechanism sits alongside a separate governance feature. The Buyer Agent now includes a vendor approval gate that connects to the IAB Diligence Platform, the compliance infrastructure IAB Tech Lab launched in 2024 in partnership with SafeGuard Privacy. When enabled, unapproved vendors are filtered out before an agent ever processes them, and the same verification check runs again immediately before a transaction is finalized. According to IAB Tech Lab, the check remains off by default and stays invisible when no verification key is present, meaning the enforcement layer activates only for organizations that specifically require it. The design intends to give regulated advertisers a governance boundary that respects agent autonomy without becoming a mandatory constraint on organizations that do not need it.

Making audiences machine-readable

The third pillar of the release focuses on Agentic Audiences, the component intended to make audience targeting usable inside the existing programmatic workflow rather than as a standalone feature. IAB Tech Lab states that the Agentic Audience Extension in OpenRTB now allows audience vector embeddings to travel inside the bidstream itself, while a supporting module built for Prebid makes that field operational on the seller side.

The release also integrates contextual enrichment technology from Mixpeek, a data management platform that donated an earlier taxonomy mapping tool to IAB Tech Lab in February. That donated technology matches contextual signals to the existing IAB Content Taxonomy standards already used throughout the programmatic supply chain, giving buyer and seller agents a shared, standards-based vocabulary for describing and matching audience segments.

IAB Tech Lab frames the audience work as a distinction between systems that merely locate an approximate match and systems that can produce a verifiable, taxonomy-backed audience plan a media buyer would actually commit budget against. Whether that framing holds up depends on how quickly buy-side and sell-side platforms adopt the extension, since a bidstream field is only useful once both sides of a transaction read and honor it.

Context: why this release differs from AAMP 2.0

AAMP's evolution has moved in stages that PPC Land has tracked since the framework's earliest components emerged in late 2025. IAB Tech Lab formally named the umbrella initiative on February 26, 2026, consolidating separate protocol pieces that had previously generated coverage and industry debate in isolation, without a shared framework tying them together. That consolidation followed a period of rapid, sometimes uncoordinated standards development: LiveRamp donated its User Context Protocol in November 2025, the Agentic RTB Framework reached version 1.0 for public comment that same month, and IAB Tech Lab published a comprehensive agentic roadmap in January extending OpenRTB, AdCOM, VAST, and OpenDirect with modern execution layers.

Version 2.0 of AAMP, released in April, focused on giving agents a broader transaction vocabulary. Version 2.3 does something structurally different: it does not add new transaction types so much as it hardens the ones already defined, addressing the operational objections that stood between a working demo and something an enterprise compliance team would sign off on. The distinction matters for a marketing technology community that has, over the past several months, watched a wide range of platforms race to announce agentic capabilities without always specifying how those capabilities would be governed once real budgets moved through them.

That governance gap has been a recurring theme across the ecosystem IAB Tech Lab is trying to organize. The organization's Agent Registry, which reached 10 active entries in March, already listed a Mixpeek MCP Server as a data management platform entry, alongside Amazon and other early participants. The registry exists specifically because, in an environment where buyer agents negotiate autonomously with seller agents, verifying a counterparty's legitimacy without an independent record becomes structurally difficult. AAMP 2.3's vendor approval gate extends that same logic from registry-level discovery down into the transaction layer itself.

The pricing provenance fix also lands inside a broader industry conversation about transparency that predates AAMP entirely. IAB Tech Lab publicly challenged Prebid's handling of transaction identifiers in August 2025, arguing that changes to the header-bidding wrapper's approach undermined cross-exchange visibility that the OpenRTB specification was designed to preserve. CPM hallucination is a related but distinct problem: rather than obscuring an existing price across the supply chain, an agent lacking real data could simply generate one. Solving that failure mode is a precondition for any protocol - AAMP or otherwise - to be trusted with programmatic spend at scale.

The audience embedding work builds directly on Mixpeek's earlier contribution to IAB Tech Lab. PPC Land reported in April that Mixpeek had donated an open-source taxonomy mapping tool in February, reducing migration work between Content Taxonomy versions from weeks of manual effort to seconds, using a combination of TF-IDF, BM25, KNN, and large language model re-ranking methods. AAMP 2.3 folds that same contextual enrichment capability into the audience layer, giving the framework a consistent taxonomy backbone across both content classification and audience targeting.

Amazon Bedrock AgentCore's role in this release also reflects a pattern PPC Land has documented across multiple platforms this year. Warner Bros. Discovery named AgentCore among six specific AWS services powering its unified advertising planning platform, describing it as the scalable infrastructure used to build, connect, and optimize its AI agents. Amazon's own advertising business has leaned on the same underlying Bedrock infrastructure, with Bedrock customer spend growing 170 percent quarter over quarter as of the company's most recent disclosures. AAMP choosing that same runtime as a deployment target signals a degree of infrastructure convergence between IAB Tech Lab's open framework and the commercial cloud environments where large advertisers and publishers already operate.

What the release does not resolve

IAB Tech Lab is explicit that AAMP 2.3 remains a maturity release built on existing foundations rather than a reinvention of the underlying protocol stack. The organization states that further use cases remain to be addressed, particularly ones extending beyond order creation and deal execution into outcomes that agents are capable of generating on their own. IAB Tech Lab has also confirmed that AAMP 3.0 is already underway, though no date has been specified for that next release.

For marketing professionals, the practical takeaway is narrower than the framework's ambitions. AAMP 2.3 does not resolve the adoption question that has followed agentic advertising standards throughout 2026: whether enough buy-side and sell-side platforms integrate the same protocol to make interoperability meaningful, rather than each major platform building a parallel, proprietary version of the same idea. Google, Amazon, The Trade Desk, and Microsoft have not been uniform in their support for competing agentic frameworks such as AdCP, and AAMP faces the same open question about how broadly its enterprise, trust, and audience layers get adopted outside the organizations already contributing to its GitHub repository.

Why marketers should pay attention

Programmatic buyers and sellers evaluating whether to pilot agentic infrastructure now have a more concrete basis for that evaluation than existed with AAMP 2.0. The pricing provenance field, in particular, addresses a concern that would otherwise disqualify agentic buying from handling real spend: without a way to trace where a negotiated price came from, an agent's output is not meaningfully different from a black box. Making that provenance auditable gives compliance and finance teams something to check against, rather than asking them to trust an agent's stated price on faith.

The vendor approval gate serves a similar function for governance teams specifically. Because the check is off by default and invisible without a verification key, organizations that do not need the added compliance layer are not forced to carry its overhead, while regulated advertisers gain an enforcement mechanism that did not previously exist inside the AAMP transaction flow.

The audience embedding work is likely to matter most to publishers and supply-side platforms integrating the Prebid module, since it determines whether audience targeting inside an agentic transaction produces a taxonomy-backed match both sides can verify, or an approximate one that a media buyer has no way to confirm. As the framework moves toward AAMP 3.0, the industry's attention is likely to remain fixed on adoption breadth rather than feature depth, since a protocol with strong governance and weak participation solves a narrower problem than one built with wide buy-in from the start.

Timeline

Summary

Who: IAB Tech Lab, the technical standards body for digital advertising, released the update. Shailley Singh, the organization's COO and EVP of Product, authored the announcement.

What: AAMP 2.3, the latest version of the Agentic Advertising Management Protocols, adds enterprise-grade deployment support through Amazon Bedrock AgentCore Runtime, a pricing provenance field that prevents AI buying agents from fabricating bid prices, a vendor approval gate connected to the IAB Diligence Platform and SafeGuard Privacy, and audience vector embeddings inside the OpenRTB bidstream supported by a Prebid module.

When: IAB Tech Lab announced the release today, roughly three months after AAMP 2.0 launched in April 2026.

Where: The framework operates globally across programmatic advertising infrastructure, with its reference code maintained in the AAMP GitHub repository under IAB Tech Lab's governance.

Why: The release addresses operational barriers that previously kept agentic advertising confined to demos and early pilots. Fixing CPM hallucination removes a failure mode that would otherwise disqualify any autonomous agent from handling genuine advertising spend, while the vendor approval gate and taxonomy-backed audience layer give compliance teams and media buyers verifiable mechanisms to trust before committing budget to agent-negotiated transactions.