IAB Tech Lab ships AAMP 3.0 with OpenProposal draft for agent-to-agent ad buying

What shipped

On September 22, 2026, IAB Tech Lab released AAMP 3.0 together with OpenProposal, a draft specification for how seller agents describe ad products in machine-readable terms and how buyer agents score those descriptions against a campaign brief. Public comment on the draft runs through October 22, 2026, according to a dev.to post examining the release.

That post also reports that IAB Australia opened a seller sandbox in the same week, letting participants rehearse discovery and proposal flows against synthetic data, with real spend explicitly excluded.

The problem being standardised

The rationale, as the post frames it, is scale: a human media team can weigh a handful of RFP responses, while an automated buyer facing thousands of them cannot rank anything unless every seller encodes inventory the same way. OpenProposal defines that shared encoding and connects it to the buying and selling plumbing Tech Lab already maintains — AdCOM, OpenDirect and the Deals API.

The draft covers the planning stage. A buyer agent starts from a brief — channel, geography, a budget range, format limits, possibly a portion of the content taxonomy — and seller agents respond with structured product definitions that the buyer scores for fit. Under AAMP 2.x, execution already spans programmatic guaranteed, preferred deals, private marketplace lines and the open exchange, all reachable through shared buyer and seller SDKs.

The missing handle

The post's sharpest criticism lands one step after the comparison. When the accepted path is programmatic, execution is still an OpenRTB bid request and response on whatever exchange is already running. Nothing in that message pair identifies which proposal won: there is no proposal identifier, no pointer to the brief version used during scoring, and nothing that ensures the video object in the request preserves the duration and placement semantics the comparison relied on.

The author contrasts this with neighbouring protocols. AdCP agrees on a version inside every message and surfaces a typed error on mismatch, and AAMP 2.3 added trust verification on price-sensitive paths, yet OpenRTB still pins its dated snapshot in onboarding paperwork rather than inside the message itself.

A concrete failure mode: video duration

The post illustrates the gap with connected TV. OpenRTB 2.6 introduced rqddurs, an array of exact acceptable durations in seconds that cannot be combined with the older minduration and maxduration range; sports and news pods depend on it because a 29-second spot in a 30-second slot leaves the slot unfilled.

Consider a brief that asks for 15- and 30-second placements. The proposal comparison scores a seller product against precisely those lengths under AdCOM placement semantics. But if the buyer agent then assembles the auction call from a template that still carries a 5-to-60-second duration range, a DSP can legitimately return 6-, 12- or 45-second creatives — each one violating what the brief and the proposal already agreed, while downstream quartile and podding logic still assumes the tighter contract. The fix the post proposes is to swap the range for the exact accepted set, rqddurs values of 15 and 30, and drop the range fields entirely.

Validating the auction hop

The post also describes rtblint, a bid request and response validator that is independent of IAB Tech Lab and required by no specification. Its broader lesson: validating the OpenProposal response against a schema accomplishes nothing if the forwarded impression object was written separately. The author recommends pinning validation to the OpenRTB snapshot version the exchange actually runs, checking AdCOM placement fields such as plcmt against what was scored during planning, and, for lines that resolve to VAST, running the resulting tag through dedicated VAST inspection tools.

Why it matters

OpenProposal automates the part of media buying that previously required humans to read proposals, and it deliberately stops at the auction boundary. Unless a later revision adds a bridge between the accepted proposal and the bid request, planning and execution remain two separate contracts, and constraints agreed during comparison can silently disappear at the auction hop. The comment window closing on October 22 is the natural moment to argue for that bridge. Until one exists, teams building agent pipelines should hold every forwarded bid request to its own validation, as a contract distinct from the proposal that selected it.