Agent registries can help establish who is operating, authentication can help establish who is making a request, and delegated permissions can establish what an agent is allowed to do. None of those facts, standing alone, establishes how independently that agent makes economic decisions.
A trustworthy name is not a theory of the decision
Picture three advertisers commissioning AI agents to compare the same set of media proposals. Each agent appears under a distinct registered identity. Each can authenticate to a trading partner. Each has permission to evaluate inventory within a separate customer's budget and rules.
That is meaningful progress. The publisher can distinguish counterparties. The advertiser can restrict authority. The intermediary has a stronger basis for logging, investigating, or rejecting an action. But what should an observer conclude if all three agents repeatedly choose the same package?
Perhaps all three follow a single centrally managed optimization policy. Perhaps they license the same underlying model but use different advertiser objectives, campaign constraints, first-party information, and instructions. Or perhaps different suppliers' systems independently converge on the same attractive deal. Identical choices are compatible with all three stories. Different identities settle none of them.
The distinction is not between trustworthy standards and untrustworthy decisions. It is between different questions that require different kinds of evidence.
Why the distinction matters
As registries, authentication and delegation mature, it will be tempting to read them as a headcount of independent actors: forty registered, verified agents bid on this inventory, so the inventory faced forty independent bidders. The inference is understandable, and it is the one this piece cautions against. It is a foreseeable misreading, not a claim that any particular party has made it.
The cost of the error runs in both directions. A publisher, advertiser or regulator who treats verified identities as independent competitors may overstate how competitive a market is, if the agents' choices are in fact closely correlated. One who treats a shared supplier as evidence of coordination may cast suspicion on agents that are in fact differentiated. Identity records, taken alone, cannot say which situation applies.
Competition policy has long assumed that firms decide independently, and recent academic work asks whether that assumption survives when firms delegate to the same or similar AI agents.[1]
What the advertising standards actually contribute
IAB Tech Lab's Agentic Advertising Management Protocols (AAMP) are an umbrella initiative designed to connect agentic workflows with established advertising standards. The Tech Lab lists AAMP v3.0, announced September 22, 2026, as its latest release. In IAB Tech Lab's announcement, WPP Media executive Michael Palmer described AAMP 3.0 as standardizing agent authentication and authorization protocols, which is precisely the layer this piece distinguishes from decision independence.[2] The accompanying OpenProposal specification addresses the path from buyer brief and publisher offering to proposal comparison and a possible buy commitment. It is a draft under public comment (extended on October 23 to November 10, 2026), not a ratified guarantee of adoption or of independently formed decisions.[2] [3] [4]
The IAB Tech Lab Agent Registry, live since March 2026, serves a different role. It lets organizations describe registered agents, purposes, capabilities, categories, maturity and connection methods. The Tech Lab says registered companies are associated with a GPP identifier and that entries are reviewed for necessary validation information. Importantly, its launch materials list further checks—such as verifying that a seller agent represents a particular property—as planned future capabilities. A listing, registration review, or verified organizational identifier should not silently become an attestation of the software's internal decision policy.[5] [6]
AAMP's open-source buyer and seller agent implementations address practical execution too. Its July v2.3 update described interoperability, OAuth-based credential changes, pricing provenance and governance controls. The Tech Lab also explains that at least one vendor-approval gate in its reference implementation is off by default unless configured. The presence of an available control should not be confused with universal enforcement by all adopters.[7]
The Agentic Real Time Framework (ARTF) v1.0 is a finalized specification for containerized services operating within a host platform and interacting with bidstream processing. IAB Tech Lab describes host-side deployment, declared capabilities, interfaces and protected mutation of bidstream messages; it has separately reported production use by implementers. Those are meaningful interoperability and execution claims. ARTF does not purport to certify that separate agent services have independently formed economic preferences.[8] [9]
And OpenRTB 2.x remains the industry protocol family for bid requests and responses. A protocol can standardize the meaning and exchange of events without guaranteeing what a particular operator retained, which upstream choices were visible, or why an agent selected its bid.[10]
None of these standards is deficient for not solving a different problem. Identity, interoperability, authorization and auditability are valuable in their own right. They can also make later verification easier—provided the relevant evidence exists and may lawfully be examined.
Six layers that should not be collapsed
When somebody says that two agents are "independent," it is worth asking independent in what sense?
| Layer | What a credible claim could establish | What it does not establish by itself |
|---|---|---|
| Legal entity | Two companies or principals are legally distinct. | Distinct software suppliers, controllers or decision policies. |
| Agent identity | Two named services or instances are discoverable and attributable at a stated identity layer. | Two independently controlled optimizers. |
| Authenticated action | A request is credibly bound to a credential or authenticated client. | That the action expresses a new, independent economic judgment. |
| Delegated authority | The principal authorized particular actions within defined limits. | Which internal reasoning or choice process produced an authorized action. |
| Technology and control | Which service providers, versions or policy operators sit behind an agent, where reliable records exist. | Equivalent behavior across customers, or unlawful coordination. |
| Observed decisions (the evidence layer) | Recorded actions resemble or differ from one another within a defined scope and context. | Their unseen causes, or a universal count of independent minds. |
The distinction between authentication and delegation is not invented for advertising. OAuth 2.0 describes limited access for an application acting on behalf of a resource owner or itself. Additional mechanisms, such as certificate-bound tokens, can strengthen proof that the presenter is the authorized client. They protect the right to act; they do not certify the provenance or independence of an economic preference.[11] [12]
Nor does a common technology supplier settle the opposite conclusion. The OECD's November 2025 examination of AI and downstream competition emphasizes both possible market-opening benefits and risks involving inputs, differentiation and common dependencies. The effect is contextual. Shared infrastructure may make independent market access cheaper; separate vendors may nevertheless produce similar choices.[13]
The hard cases are more instructive than the obvious ones
Same supplier, different judgment. A shared agent platform evaluates a luxury advertiser's brand-safety requirements and a performance advertiser's acquisition target. Different approved budgets, objectives and information yield different plans. The common supplier is important provenance, but not proof of common economic judgment.
Different credentials, common policy. Three registered agents connect through separate authorized accounts, yet all use a central service whose outputs are determined by a single policy with nearly identical inputs. The separate credentials are genuine; the degree of policy differentiation remains a separate empirical question.
Different systems, same good answer. Three agents built on unrelated technology independently favor the same media placement because the price, audience fit and expected returns are unusually favorable. Their agreement is not evidence of a hidden shared controller, still less collusion.
All three are hypothetical, not findings about any company or marketplace. They show why a registry entry, an architectural diagram, or a correlation chart cannot alone support a stronger diagnosis.
There is also a definitional limit. One organization can operate several truly different policies. Two organizations can independently sample actions from identically specified policies. Whether a market is evaluating supplier diversity, control independence, policy differentiation or dependence between observed actions changes both the evidence needed and the interpretation. There is no single universal "independent agent" badge that resolves all of those targets.
What a serious question to an operator sounds like
Before calling a market's decisions independent—or concentrated—an operator can ask four practical questions.
Which actor is represented? Is the record naming a legal advertiser, an agency, a DSP seat, a hosted agent, a service account, a credential, or the organization controlling a policy? These units may not map one-to-one.
What was authorized and what actually happened? A valid permission establishes a boundary. An execution record documents an action. Neither necessarily captures the alternatives the system considered or distinguishes agent selection from a human-approved instruction.
What decision evidence exists? Are the opportunity, contemporaneous commercial context, selected action and relevant provenance sufficiently documented and permissioned to support the specific inference? A record of an accepted or winning action cannot describe decisions excluded upstream from observation. For that separate issue, see our article on auction transparency and decision evidence.
Which explanations remain live? Shared objectives, common market conditions, synchronized data, platform routing, shared suppliers and genuine policy convergence can produce superficially similar outcomes. An inquiry that cannot distinguish them should state its limit, not assign a hidden cause.
What would move each answer? Different layers call for different evidence. Legal distinctness rests on corporate records. Identity and authority rest on registry entries, credentials and delegation logs. Supplier and control questions rest on disclosures or attestations of which providers, versions and policy operators are involved. Questions about decisions themselves rest on records of the opportunity, the commercial context, the action taken and, where available, the alternatives considered, across enough comparable opportunities to separate pattern from coincidence. Recent academic work also points toward behavioral evidence: a 2026 ICML position paper reports that, in simulated pricing games, a model's reasoning traces did not track its collusive behavior, and argues that certification should rest on outcomes observed in representative situations rather than on stated intent or reasoning.[1] What is obtainable will vary by marketplace.
This is not an argument that every commercial decision must expose private model weights, prompts, customer data or confidential source code. Privacy, intellectual property, competition sensitivity and data-access rights remain real constraints. Nor should an agent be compelled to reveal restricted internal reasoning simply because a counterparty wants a comforting label. A defensible claim must be calibrated to the available, authorized evidence; some questions will remain unanswered.
The strongest argument for the standards is also the case for precision
The case for agent identity is strong even if two registered systems later prove to use closely related tools. Verified identity supports dispute resolution and accountability. Delegated authority limits what software may do. Common interfaces reduce integration costs. Consistent records may one day create opportunities for more rigorous, independently reviewed inquiry.
But capability is not implementation, implementation is not observation, and observation is not causal explanation. An announced specification does not mean universal deployment. A permissioned transaction does not mean the agent autonomously chose the underlying commercial objective. Behavioral similarity does not establish misconduct or the existence of a shared model.
For technical standards developers, the opportunity is to preserve clear provenance and authority boundaries without claiming to solve every measurement problem. For marketplace operators, it is to know which identities, decisions and exclusions their records can actually describe. For advertisers, it is to distinguish who was allowed to spend from why a particular spend decision was made.
Exchange Optics is examining the broader problem of independent decision evidence in automated marketplaces. That is not a claim of IAB Tech Lab endorsement, of live-market results, or of universal measurement capability. Where evidence is inadequate, the responsible answer is unresolved.
Agent identity tells us whose system appears at the door. Authorization tells us whether it may enter. Interoperability helps it do business once inside. Whether the marketplace contains meaningfully independent decisions is another question—and its answer must be earned with evidence.
References
[1]Riemer, M. et al., Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions, Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:171880-171907 (https://proceedings.mlr.press/v306/riemer26a.html) (2026; accessed October 31, 2026).
[2]IAB Tech Lab, AAMP 3.0 and OpenProposal announcement (https://iabtechlab.com/press-releases/iab-tech-lab-introduces-aamp-3-0-with-openproposal/) (September 22, 2026; live page records extension to November 10, 2026; accessed October 31).
[3]IAB Tech Lab, AAMP initiative and releases (https://iabtechlab.com/standards/aamp-agentic-advertising-management-protocols/) (last updated September 22, 2026; accessed October 31, 2026).
[4]IAB Tech Lab, OpenProposal public-comment specification repository (https://github.com/IABTechLab/OpenProposal) (v3.0-draft-1 announced September 22, 2026; repository README says not ratified; accessed October 31).
[5]IAB Tech Lab, Introducing the Agent Registry (https://iabtechlab.com/introducing-the-iab-tech-lab-agent-registry/) (March 2026; features and roadmap; accessed October 31).
[6]IAB Tech Lab, 2026 Roadmap (https://iabtechlab.com/navigating-the-agentic-frontier-the-iab-tech-lab-2026-roadmap/) (2026; identifies March 3 registry go-live; accessed October 31).
[7]IAB Tech Lab, Agentic Advertising Matures with AAMP 2.3 (https://iabtechlab.com/agentic-advertising-matures-with-aamp-2-3/) (July 2026; implementation details and default-off gate; accessed October 31).
[8]IAB Tech Lab, ARTF standard page (https://iabtechlab.com/standards/artf/) (updated August 12, 2026; states finalized v1.0; accessed October 31).
[9]IAB Tech Lab, finalized ARTF v1.0 specification (https://github.com/IABTechLab/agentic-real-time-framework/blob/main/Agentic_Real_Time_Framework_Version_1_0_FINAL.md) (specification front matter: released November 12, 2025; see also finalization announcement (https://iabtechlab.com/a-milestone-for-the-agentic-advertising-ecosystem/); accessed October 31, 2026).
[10]IAB Tech Lab, OpenRTB standards hub (https://iabtechlab.com/standards/openrtb/) (version history and OpenRTB 2.x links; accessed October 31, 2026).
[11]IETF, RFC 6749, The OAuth 2.0 Authorization Framework (https://datatracker.ietf.org/doc/html/rfc6749) (October 2012).
[12]IETF, RFC 8705, OAuth 2.0 Mutual-TLS Client Authentication and Certificate-Bound Access Tokens (https://datatracker.ietf.org/doc/html/rfc8705) (February 2020).
[13]OECD, Artificial intelligence and competitive dynamics in downstream markets (https://www.oecd.org/en/publications/artificial-intelligence-and-competitive-dynamics-in-downstream-markets_ccf0624a-en.html) (November 14, 2025).
Standards and cited implementation information were reviewed as of October 31, 2026. Identity, authorization and interoperability answer different questions from effective economic decision diversity.