16 Oct 2026 · 7 min read
What Do AAMP's Buyer-Agent and Seller-Agent Reference Implementations Do, and How Should Teams Use Them?
AAMP's buyer-agent and seller-agent are open-source reference implementations from IAB Tech Lab. They show, in working code, how demand-side and supply-side agents can price, negotiate and book programmatic direct deals using OpenDirect 2.1 and the IAB Deals API. They are a worked specification, not a product to deploy unchanged, and their most useful lesson is where they place human approval.
TL;DR. According to the AAMP hub repository, the buyer agent demonstrates demand-side agent logic and decision workflows, and the seller agent demonstrates supply-side coordination and marketplace interaction. The buyer agent turns a campaign brief into negotiated bookings and holds them for human approval unless auto-approve is on. The seller agent publishes a tiered media kit, evaluates proposals and distributes deals; its approval gate is off by default. Both include clients written for agentic-direct, the AAMP component that implements OpenDirect 2.1 over MCP and A2A. Before building on either, a team should pin a tagged release, read the changelog and replace the demo data layer.
What are AAMP's buyer and seller reference agents?
AAMP, the Agentic Advertising Management Protocols, is IAB Tech Lab's open-source framework for agentic advertising, announced in early 2026 and published as separately versioned repositories. IAB Tech Lab's AAMP page calls buyer-agent and seller-agent the "Buyer and Seller Agents SDKs", and its Agent Registry announcement says AAMP "provides reference implementations for creating Agentic Advertising Agents that follow existing standards".
Both are Python applications built on CrewAI and licensed under Apache 2.0. At the time of writing, the latest tagged release of each is v2.4.2, published on 18 August 2026. The buyer agent's documentation calls it an alpha release; the seller agent's README marks its second version as feature complete.
What does the buyer agent demonstrate?
The buyer agent demonstrates a decision workflow that starts from a campaign brief. According to its documentation, the booking flow validates the brief, splits the budget across branding, CTV, mobile and performance channels, researches seller inventory for each channel in parallel, optionally negotiates, consolidates recommendations, pauses for human approval and then books.
Negotiation is a pluggable strategy. The one strategy marked available, SimpleThresholdStrategy, opens at a target CPM, concedes a fixed step each round, accepts any price at or below a maximum CPM, and walks away after a set number of rounds or when the seller stops moving. Booking follows a quote-then-book pattern through the IAB Deals API v1.0, and the buyer can disclose its identity in tiers (public, seat, agency, advertiser) to receive different pricing.
What does the seller agent demonstrate?
The seller agent demonstrates how a publisher or SSP might expose inventory to agents and handle a buyer agent's requests. Its README describes automated deal negotiation, booking and distribution using OpenDirect 2.1, the Deals API v1.0 and sellers.json.
That covers a media kit with public and authenticated views, identity-based price tiers with floors, an agent that reviews proposals and recommends accept, counter or reject, deal ID generation, and distribution by Deals API push, ad server or SSP. Orders move through a state machine from draft to booked, delivering and complete, and deal responses carry OpenRTB supply chain (schain) information. MCP is the primary interface.
How do the reference agents use agentic-direct?
The AAMP hub repository describes agentic-direct as providing "the core communication framework and orchestration layer for agent interaction". The agentic-direct README describes a schema-driven implementation of the A2A protocol with MCP integration for OpenDirect v2.1: it generates MCP tools from an OpenDirect schema and serves separate buyer and seller agents. As a reference server, it is there to show the message flow, so teams connect their own systems behind it.
The reference agents build on that layer. The buyer agent's MCP, A2A and unified clients are each described in the source as clients for the "IAB agentic-direct server", addressed through an IAB_SERVER_URL setting. The seller agent's OpenDirect 2.1 client connects to the same server. In short, agentic-direct supplies the shared OpenDirect vocabulary and transport, and each agent adds its side's decision logic. AdCP, the separate Ad Context Protocol from AgenticAdvertising.org, defines its own task set, so teams supporting both map each separately.
Where do human approval and limits appear in the code?
In the buyer agent, a booking request carries an auto_approve flag that defaults to false. With that default, the flow stops at awaiting_approval and books only what a person approves. Before booking, the flow enforces a spend ceiling against the campaign budget, and the negotiation strategy's maximum CPM and round limit cap price and duration.
In the seller agent, the approval gate is opt-in: APPROVAL_GATE_ENABLED is false in the example configuration. When enabled, the proposal_decision gate pauses after the review agent recommends. An operator approves, rejects or counters with modified terms, records who decided and why, and resumes the flow; pending approvals expire after 24 hours by default. A value threshold, approval_required_above_value, forces human approval above a set deal value, but defaults to zero, which disables it.
What should a team check before building on reference code?
Pin a tagged release. Main moves ahead of releases. At the time of writing, the buyer README on main already points to a v2.5.0 upgrade guide, while the latest tag is v2.4.2.
Read the changelog. Reference code improves release by release. Before upgrading or forking, read the changelog for the tag you are on so you know which behaviours have changed since.
Check which interfaces each release implements. MCP is the seller agent's primary interface, and its documentation sets out which other interfaces are available in the current release. Confirm the interface your counterparties use before you build on it.
Replace the demo data layer. The seller agent's production guide states that the demo uses CSV files and in-memory SQLite, and explains how to connect a real ad server and database.
Set the approval policy first. Enable the seller gate, set a value threshold, leave auto_approve false and name the approver.
IAB Tech Lab's Agent Registry announcement says that "as with all things AAMP and Agentic we are still in the 'test and learn' phase". The team's job is to know which version it took and what it changed. Building on the reference agents confers no certification, because no AAMP compliance certification exists.
What do the reference agents leave to the deploying team?
Each reference agent keeps its own records. The buyer stores deals, negotiation history and an order audit log; the seller keeps an order audit log and an event log. In a delivery dispute, what matters is one record of agreed terms that both sides hold, and that sits with the parties or a marketplace.
Alkimi's neutral marketplace is one place that record can live. It works within IAB Tech Lab's AAMP framework, its agent is listed in the IAB Tech Lab Agent Registry, and it is being built to be AdCP compatible. Agreed deals are written to a shared DealSheet both parties hold, with humans keeping approval rights at the deal level.
Frequently asked questions
Are AAMP's buyer and seller agents production software?
They are reference implementations. The buyer agent's documentation calls it an alpha release, and the seller agent's production guide explains how to replace its demo data layer.
Do the AAMP reference agents require human approval?
The buyer agent holds recommendations for approval by default. The seller agent's approval gate is opt-in and off by default.
What does agentic-direct do for the buyer and seller agents?
agentic-direct is AAMP's communication and orchestration layer. It implements OpenDirect 2.1 over A2A and MCP, and both reference agents include clients written to connect to it.
Does building on the reference agents make a product AAMP compliant?
No. There is no AAMP compliance certification. IAB Tech Lab runs an Agent Registry where companies list their agents, which is a listing, not a certification.