The Pievra
Protocol Briefing

What AdCP Is

AdCP is an open technical standard that enables AI agents to communicate with advertising platforms through a unified interface — and, as it evolves, directly with each other. It is the agentic successor to OpenRTB: where OpenRTB standardised real-time bidding transactions, AdCP standardises agentic advertising workflows across the entire campaign lifecycle.

AdCP is built on top of Anthropic's MCP and Google's A2A as its transport layers. It defines domain-specific advertising tasks and schemas that run over these general-purpose agent communication protocols. MCP and A2A are the infrastructure; AdCP defines the advertising vocabulary and rules running over that infrastructure.

The Four Protocol Domains (v2.5 — current production)

• Signals Activation Protocol: Agents discover and activate audience signals — identity, contextual, geographic, temporal — across connected platforms using natural language queries.
• Media Buy Protocol: Standardises how agents execute and optimise media purchases across all compatible platforms from one interface.
• Creative Protocol: Enables agents to build, optimise, and deploy creative assets. AI-generated creative integrates into the workflow.
• Curation Protocol: Marked "Coming Q2 2025" at launch; in active development in v3.0. Allows inventory curation with activated signals.

How It Works in Practice

A buyer's AI assistant receives a natural language brief: "Reach eco-conscious car buyers on CTV in the US this week, $10K budget." The agent translates this into structured AdCP requests, queries compatible Seller Agents via MCP or A2A, receives standardised responses, and executes the buy — without custom integrations per platform. Prebid's stewardship of the Seller Agents project is the mechanism ensuring broad publisher adoption without fragmentation.

AdCP's governance layer adds check_governance validation — verifying budget limits, brand safety, and compliance before any campaign launches, with a complete audit trail of every agent decision.

What MCP Is and Is Not

MCP is a universal context exchange and tool-connection standard — the infrastructure layer on which advertising protocols including AdCP, ARTF, and Agentic Audiences are built. It defines how AI agents connect to external tools, data sources, APIs, and resources. It is not an advertising protocol. It is the plumbing beneath the advertising protocols.

In advertising: MCP is what allows a campaign agent to call a DMP for audience data, retrieve brand guidelines, check compliance rules before bidding, and retain cross-campaign memory — all through one standardised interface without custom integrations per data source.

Specification Milestones

• November 2024 (launch): Basic client-server, tool calling, resource access, prompt templates. Synchronous operations only.
• March 26 2025: OAuth 2.1, Streamable HTTP (replaces HTTP+SSE), JSON-RPC batching, Tool Annotations. OpenAI adopts across all products. Microsoft launches Playwright-MCP.
• June 18 2025: Structured tool outputs, server-initiated user interactions (elicitation), improved security, formal governance model.
• September 2025: MCP Registry preview — public catalog and API for server discovery.
• November 25 2025 (current production): Asynchronous task execution, .well-known server identity discovery, enterprise horizontal scaling improvements, long-running workflow governance.
• December 2025: Donated to Linux Foundation AAIF — vendor-neutral governance.

What Agentic Audiences (UCP) Is

Agentic Audiences is the data plane of agentic advertising — the standard defining how agents exchange user signals. It is not a campaign execution protocol. It defines how buyer agents, seller agents, and measurement agents share what they know about a user's intent, identity, and advertising response history without exchanging raw personal data.

The technical innovation: embeddings — compact, learned vector representations of 256 to 1,024 dimensions that encode complex user signals in a format that is simultaneously privacy-preserving, fast, and interoperable. Instead of passing thousands of raw data points, agents exchange dense vectors that capture semantic meaning. Similar user intents produce similar vectors. An agent can compare and act on these without ever resolving the underlying identity.

Three Signal Types

• Identity signals: Hashed identifiers, behavioural segments, probabilistic matches — in compressed vector form, no raw PII
• Contextual signals: What the user is doing right now — content category, time of day, device context
• Reinforcement signals: How this user or users like them have previously responded to advertising — enabling real-time model learning

AdCP defines how agents act on the control plane. Agentic Audiences defines the data plane — the signals that inform those actions. They are designed to work together. IAB Tech Lab explicitly positions UCP as the data layer that extends and enriches AdCP-based agent workflows.

The AAMP Context

On February 26, 2026, IAB Tech Lab formally named its umbrella initiative AAMP — Agentic Advertising Management Protocols. Agentic Audiences is one of seven protocol components within AAMP, alongside ARTF, Agentic Ad Objects (AdCOM), Agentic Direct, Agentic Deals, and two components in development. Understanding UCP requires understanding it as part of a coordinated infrastructure — not a standalone protocol.

What ARTF Is

ARTF is a containerisation architecture for programmatic advertising. It allows technologies — DSPs, SSPs, enrichment partners, measurement solutions, AI agents — to operate within the same virtual environment rather than communicating across separate networks via API calls. The result: auction-speed agentic decisioning that was previously impossible within RTB latency constraints.

ARTF is explicitly different from AdCP. AdCP defines how agents communicate in the planning and buying layer above the auction. ARTF defines how agents participate in real-time bidding execution at the infrastructure level, within the auction. IAB Tech Lab describes them as complementary rather than competing.

How Containerisation Works

ARTF packages agent or technology code into a software container that runs inside a partner's infrastructure. A DSP's bidding logic, a fraud detection agent, and a viewability measurement service can all operate in the same virtual machine — communicating at memory speed rather than network speed. This eliminates the network hops that create the 600–800ms round-trip time of current OpenRTB implementations.

The gRPC service definition: RTBExtensionPoint.GetMutations(RTBRequest) — an agent receives a bid request, proposes mutations (segment activations, deal adjustments, bid shading, viewability additions), and returns them within the latency budget. A working Go reference implementation is publicly available on IAB Tech Lab's GitHub.

What A2A Is

A2A is a universal agent-to-agent communication protocol — the standard enabling any two AI agents, regardless of builder or framework, to discover each other, authenticate securely, and collaborate. It is the "HTTP for AI agents." A2A is complementary to MCP, not competing: MCP connects an agent to its tools (agent-to-tool); A2A connects agents to other agents (agent-to-agent). Google's documentation explicitly recommends using them together.

Core Architecture

Every A2A agent publishes an Agent Card — a JSON metadata document at /.well-known/agent.json advertising its identity, capabilities, service endpoint, and authentication requirements. Agent Cards enable discovery without a prior relationship. A buying agent queries a registry, receives Agent Cards for matching sellers, and initiates direct authenticated communication without an exchange.

Tasks are the unit of work: structured objects with a defined lifecycle (submitted → working → input-required → completed / failed). Long-running campaigns maintain state across the lifecycle, with agents pushing status updates to registered notification endpoints.

Version History

• April 9 2025 (v0.1): Core spec — Agent Cards, task management, basic auth. 50+ partners.
• May 20 2025 (v0.2 / Google I/O): Stateless interactions, OpenAPI-like auth schemas, official Python SDK. Microsoft Azure and SAP announce support.
• June 2025: Donated to Linux Foundation — vendor-neutral governance.
• July 31 2025 (v0.3): gRPC support, signed Agent Cards (cryptographic identity verification), extended Python SDK, native ADK integration, Agent Engine deployment support.