The IAB Tech Lab published the AAMP Seller Agent SDK on GitHub. Open source. Working reference implementation. Docker, CloudFormation, Terraform, Kubernetes deployment templates. Redis, Postgres, or SQLite backends. Auto-scaling, VPC networking, audit logging.
For the first time, a publisher CTO can read the code that defines what a sell-side agent stack looks like. Not a whitepaper. Not a webinar. A repository.
I read it with one question: how far is D/Cipher from this architecture?
The answer is closer than I expected — and further than it should be.
The Three-Tier Hierarchy
The Seller Agent SDK implements a three-tier agent hierarchy that mirrors the buyer side. Both sides of the transaction speak the same language — AdCOM schemas, OpenDirect 2.1 deal objects, OpenRTB supply chain protocols.
Strategy. Revenue targets. Brand constraints. Which buyers to prioritize, which to gate. The decisions a VP of Revenue makes — encoded as agent policy.
CTV. Display. Video. Mobile. Native. Each specialist translates portfolio strategy into channel-specific execution — rate cards, floor pricing, audience packages.
Bounded tasks. Pricing calculations. Avails lookup. Audience qualification. Deal state management. The work that should never be an LLM decision — structured, auditable, fast.
AAMP Seller Agent SDK — three-tier hierarchy. Source: IAB Tech Lab
The hierarchy matters because it separates what an LLM should decide from what it absolutely should not. Portfolio strategy involves judgment — which buyer relationships to deepen, how to balance fill rate against CPM. Functional tasks involve execution — looking up available inventory, computing a price floor, serializing a deal object. The SDK keeps LLM reasoning at the top and deterministic computation at the bottom. Smart architecture.
What the SDK Actually Contains
This is not a specification document. It is running code with production-grade infrastructure.
| Component | What It Does |
|---|---|
| Media Kit Manager | Tiered access — public discovery data vs. authenticated buyer detail. Dynamic adaptation based on buyer relationship and negotiation stage. |
| Negotiation Engine | Configurable strategies. Rate card mapping. Human-in-the-loop approval gates. Automated responses within guardrails, escalation outside them. |
| State Machine | Deal and order lifecycle management. Configurable workflow stages. Event-based notifications. Full audit trail. |
| Inventory Connectors | Google Ad Manager. FreeWheel. PubMatic. Index Exchange. Magnite. sellers.json and OpenRTB supply chain. |
| Agent Registry | Discovery endpoint. Trust signals. Capability declarations. How buyer agents find you. |
| Reference Deployment | Docker, AWS CloudFormation, Terraform, Kubernetes. Redis/Postgres/SQLite. Auto-scaling. Production-ready. |
The inventory connectors are the detail that stops you. Google Ad Manager is a supported integration. D/Cipher runs on Graphene-Selene, which uses GAM as its ad server. The SDK already speaks to the infrastructure D/Cipher sits on.
The Mapping
Here is what D/Cipher has, what the Seller Agent SDK expects, and what is missing.
| SDK Requirement | D/Cipher Today | Gap |
|---|---|---|
| Contextual intelligence | 300K+ content embeddings, offline inference, proprietary taxonomy | None |
| GAM integration | Graphene-Selene uses GAM as primary ad server | None |
| OpenRTB compatibility | Programmatic revenue flows through OpenRTB infrastructure | None |
| Media Kit endpoint | Bilateral sales decks. No machine-readable surface. | Build |
| Negotiation engine | Human sales team. No programmatic deal negotiation. | Build |
| Deal state machine | CRM-managed deals. No automated lifecycle. | Build |
| Agent Registry entry | Not registered. 10 entries, D/Cipher absent. | Register |
| MCP server | None. Spec finalizes July 28. | Build |
| Protocol vocabulary | Proprietary taxonomy. Not serialized to AdCOM/OpenDirect. | Map |
Three of nine requirements already exist inside D/Cipher. The intelligence is there. The ad server integration is there. The programmatic pipes are there. What is missing is the surface — the protocol layer that makes all of it queryable by a machine buyer.
Two Tiers of Participation
A CTO reading the SDK sees two paths, not one.
Expose D/Cipher's contextual intelligence and inventory data via MCP. Stateless protocol. OAuth 2.1 authentication — People Inc already runs Keycloak. Buyer agents can query. You exist on the agentic surface.
Deploy the three-tier hierarchy. Media Kit serialized from D/Cipher taxonomy. Negotiation engine with rate cards and approval gates. Deal lifecycle management. Agent Registry listing. Full autonomous transactability.
The minimum path is the one I estimated in The Two Muscles: three engineers, two sprints. The MCP stateless spec removes sessions, supports horizontal scaling behind a load balancer, uses OAuth 2.1 for authentication. People Inc has the IAM infrastructure. The spec is designed to be easy for exactly this kind of deployment.
The full path is new — and the SDK makes it real. The reference implementation runs on Amazon Bedrock AgentCore. It supports every deployment model People Inc's infrastructure team would recognize. The deal state machine extends OpenDirect 2.1, which D/Cipher's programmatic team already speaks. The negotiation engine's human-in-the-loop gates preserve the direct-sales relationship Roberts values. This is not a replacement for the sales team. It is a protocol surface that the sales team's existing agreements run through.
What the Field Looks Like
While the blueprint sits unread, the field fills in around it.
Adform opened 800+ capabilities via MCP on July 8 — the full FLOW platform, every stage of the advertising lifecycle, queryable by any AI agent. PubMatic Decision Fabric containerizes partner models inside the supply path at 1ms inference. LiveRamp LAB opened the identity graph to third-party agents via MCP. Amazon, Meta, Microsoft, and TikTok all run MCP servers for their ad platforms.
The AAMP Agent Registry has 10 sell-side entries: Amazon Ads, PubMatic, Equativ, Optable, Dstillery, Burt Intelligence, HyperMindZ, Mixpeek, IAB Tech Lab. D/Cipher is not among them.
78% of enterprise AI teams now run MCP-backed agents in production. 28% of Fortune 500 companies operate MCP servers. Monthly SDK downloads: ~97 million. This is not an emerging standard. This is the standard.
The Narrowing
The SDK reveals something I had not fully articulated in 48 previous posts: D/Cipher's competitive advantage does not have a format problem. It has a serialization problem.
The intelligence exists. 300,000+ content embeddings across 40 brands. Offline inference that avoids client-side latency. A proprietary contextual taxonomy that Roberts correctly calls one of publishing's strongest. That intelligence is real and valuable.
But it is locked in a format that buyer agents cannot read. The Seller Agent SDK's Media Kit component exists to solve exactly this — serialize your proprietary intelligence into a machine-readable surface with tiered access. Public data for discovery. Authenticated data for negotiation. Progressive detail as trust deepens. The format preserves competitive advantage while enabling machine discovery.
D/Cipher's contextual taxonomy, serialized as an AAMP Media Kit, would be one of the strongest sell-side intelligence surfaces in the registry. Better than most of the 10 entries that are there. The content graph across Allrecipes, People, InStyle, Better Homes & Gardens, Food & Wine — that is a differentiated inventory signal that no other publisher can match at that scale.
The blueprint is published. The integrations connect to D/Cipher's existing infrastructure. The deployment templates run on infrastructure People Inc's team already operates. The negotiation engine preserves human oversight. The trust layer protects competitive intelligence.
What remains is a decision. Not an engineering decision — a governance decision. The same one we have been watching for seven months. Someone at People Inc has to say yes. The SDK does not change what needs to happen. It removes the last ambiguity about how.
AAMP 3.0 ships in July. The MCP spec finalizes July 28. The reference implementation is live on Amazon Bedrock AgentCore. Every week, buyer agents optimize toward what responds. Defaults calcify.
The blueprint is public. The question is whether anyone at People Inc has opened the repository.