Agentic RTB Framework – Overview and Technical Details

This page summarizes key elements of the Agentic RTB Framework (ARTF) in a single place for technical readers. It is intended as a neutral, implementation-focused reference for engineers, product managers, and architects. Definitions of key terms are provided in the ARTF glossary.

1. Scope and positioning

  • Standard type. ARTF is an execution-layer framework for real-time bidding. It defines how host platforms integrate and orchestrate containerized "agent services" that can enrich, evaluate, or modify the bidstream.
  • Relationship to OpenRTB. ARTF is designed to coexist with, not replace, OpenRTB. OpenRTB continues to define the bid request and response objects; ARTF defines how services interact with and "mutate" those objects inside the auction.
  • Status. Version 1.0 of ARTF defines the initial containerized execution model and supporting interaction patterns.

At a high level, ARTF sits in the execution layer below protocol and data definitions. OpenRTB remains the bid object; ARTF standardizes how it is processed by co-located services.

graph LR Exec[Execution layer ARTF host and agents] Comm[Communication layer RPC and agent protocols] Data[Data layer OpenRTB bid objects] Exec --> Comm Comm --> Data

2. Core architecture

ARTF describes an architecture with three main elements:

  • Host platform (orchestrator). The host is the SSP, DSP, exchange, ad server, or curation platform that runs the core auction or decisioning workflow. It owns the infrastructure, controls access to data, and decides which agents to invoke and which mutations to apply.
  • Agent services. Agents are third-party or first-party services packaged as containers. Examples include identity resolution, fraud detection, contextual or audience enrichment, bidding logic, or measurement.
  • Containers. Agents are delivered as standardized, lightweight containers that can run in the host's environment (for example, in the same data center, on the same server, or in the same Kubernetes cluster). Containerization provides isolation while allowing efficient, low-latency execution.

The host is responsible for provisioning, scheduling, and monitoring agent containers, enforcing resource limits and security policies.

3. Data and security model

A central design goal of ARTF is to support collaboration on sensitive data while preserving host control:

  • Protected bidstream mutation. Agents interact with the bidstream through a standardized API that exposes only the data needed for a given task and accepts proposed changes as structured mutations.
  • Least-privilege data access. The host determines which fields of the bid request or related context are shared with each agent, applying a "least data" principle.
  • Host control of SLAs. All inputs to and outputs from agents are mediated and audited by the host, which can accept or reject proposed mutations individually.
  • Isolation and integrity. Agents run in sandboxed containers, and the specification supports mechanisms such as cryptographic signing of container images so that hosts can verify integrity before deployment.

This model is designed to reduce data leakage risks compared with architectures that require full bid requests to be sent to multiple external endpoints.

4. Latency and efficiency

In the legacy architecture, a single impression may trigger multiple sequential external calls for enrichment, identity, and fraud checks, leading to end-to-end latencies in the range of several hundred milliseconds.

By moving these functions into co-located containers and using in-process or intra-data center communication, ARTF is designed to:

  • Reduce bid request/response times by up to approximately 80%, from typical ranges of 400–800 ms down to roughly 100 ms, subject to implementation.
  • Treat the recovered time as a "first-class asset" that can be re-invested into richer, real-time enrichment and decisioning within the auction window.
  • Improve computational and energy efficiency by avoiding large volumes of redundant cross-internet traffic.

5. Interaction protocols and OpenRTB Patch

ARTF standardizes how hosts and agents communicate and how agents express changes to the bidstream:

  • gRPC and Protocol Buffers. The framework uses gRPC with Protocol Buffers for high-performance, strongly typed communication between the host and agent containers. The IAB Tech Lab provides authoritative .proto definitions for the request and response messages and services.
  • OpenRTB Patch. Agents return proposed changes as OpenRTB Patch objects, which are collections of atomic "mutations" that reference paths within the OpenRTB object and are tagged with an intent.
  • Intents. Each mutation includes an intent (for example, identity resolution, fraud detection, deal activation, or price adjustment) that describes the purpose of the change. Intents help the host sequence agents, apply policy, and audit behavior.

Instead of replacing entire bid requests or responses, agents return minimal diffs, and the host decides which mutations to apply.

6. Agent orchestration

Hosts can configure multiple agents to run on the same impression:

  • Agents may be invoked sequentially or in parallel, depending on dependencies.
  • Typical sequences might include identity enrichment, contextual enrichment, and fraud evaluation before a request is sent to buyers, or validation on bid responses.
  • The orchestrator logs which agents were invoked, what mutations they proposed, and which changes were ultimately applied.

This approach allows new capabilities to be added by deploying new containers and updating orchestration policies, without changing the core auction implementation.

7. Relationship to MCP and A2A

ARTF v1.0 also anticipates more advanced agentic scenarios:

  • Model Context Protocol (MCP). MCP is a separate, JSON-RPC-based protocol that allows AI models and tools to access external resources and services. ARTF is defined to support MCP as a complementary protocol for model-to-service and model-to-agent interactions.
  • Agent-to-Agent (A2A) communication. The specification references the possibility of agents collaborating directly, where appropriate, in addition to host-mediated service-to-service calls.

Together, these elements position ARTF as a foundational execution environment for both current systematic integrations and future autonomic, AI-driven agents in programmatic advertising.

Related pages: