ARTF, the Agentic Real Time Framework, is a technical specification from the IAB Tech Lab that allows one company to run its software inside another company's advertising auction. Rather than a supply-side platform sending a bid request out across the public internet to a fraud vendor, a data provider or a buyer's pricing model and waiting for an answer, the vendor packages its logic as a container image, hands it over, and the host runs it inside its own data centre. The container has no route to the outside world. It receives a portion of an OpenRTB bid request, proposes changes to it, and the host decides which of those changes to apply. The standard exists because a network round trip to an external service consumes more milliseconds than a real-time auction can spare, and because the alternative, giving every partner a bespoke integration, has proved slow and expensive to build.

OpenRTB is the IAB Tech Lab protocol carrying bid requests and responses between sellers and buyers. gRPC is a binary remote procedure call framework. MCP, the Model Context Protocol, is a JSON-RPC interface built for connecting language models to tools. ARTF uses all three.

How a mutation reaches the auction

The unit of work in ARTF is the mutation: a single, atomic proposed change to a bid request or bid response. The host platform, which the specification calls the orchestrator, exposes an extension point at a defined moment in the auction lifecycle. It calls each registered container over gRPC using a service named RTBExtensionPoint, whose only method, GetMutations, takes an RTBRequest and returns an RTBResponse.

The request envelope carries a lifecycle value identifying the auction stage, an identifier, the OpenRTB bid request, optionally a bid response, and tmax, the millisecond budget within which the answer must arrive including network latency. The container receives only the fields the orchestrator chooses to send. Least-data is one of the specification's five founding principles, and it instructs orchestrators to strip private and competitive signals before an agent sees anything.

Each mutation in the reply carries three fields and a payload. The intent declares why the change is wanted. The operation is add, replace or remove. The path points at the target. The payload is one of four typed structures: a list of identifiers, a deal adjustment, a bid price adjustment, or a set of metrics.

Paths are the subtle part. Rather than pointing at a literal position in a JSON document, which the agent cannot reliably know, they use semantic references drawn from OpenRTB concepts. The published example activating audience cohorts declares the intent activateSegments, an add operation, a path of /user/data/segment, and a payload of two segment strings. A longer example expires two deals on the first impression, adds two more, then replaces deal floors at 5.00 and 8.00. That indirection lets orchestrators reorganise their internal data structures without breaking every agent integration.

Nothing binds those four mutations together. The specification is explicit that each patch is atomic, must be accepted or rejected whole, and that there are no transactions and no ordering guarantees across mutations. Whether the orchestrator applies a change is entirely at its discretion, and whether it tells the agent what it decided is, in the specification's phrasing, left to the parties.

What the host demands of the container

Five requirements govern participation. Agents must transact in the core bidstream rather than sit at the edge. They must declare specific intents. They must ship as containers conforming to Open Container Initiative image and runtime standards, so they can be managed by Kubernetes, Docker Compose or Amazon Elastic Container Service. They must be fast, which the specification enforces by mandating gRPC with protobuf serialisation and recommending Rust, Go or Java. And they must operate under least privilege and least data.

The isolation rules are strict. The container runs as a non-root user, and the orchestrator will never run it as root. All network ingress and egress is prohibited except communication with the orchestrating entity, though the two parties may negotiate exceptions. Privileged access and host network or process namespaces are forbidden, unnecessary Linux capabilities are dropped, and each container is isolated and unaware of the cluster around it. Kubernetes-compatible liveness and readiness probes over HTTP are mandatory, as is Open Telemetry for metrics and distributed tracing.

Capabilities are declared in an agent manifest, a JSON structure embedded in the container image as metadata under the label com.iabtechlab.artf.descriptor, Base64 encoded and forbidden from carrying secrets. It names the agent, its vendor, its owner, the intents it supports, its minimum processor and memory requirements, and any host services it depends on by name. The specification's example manifest requests 500 millicores and 256 mebibytes and declares seven intents including bidValuation, dynamicDealCuration and audienceSegmentation. If the orchestrator cannot meet a manifest's requirements, the container should not be started.

The Go reference implementation in the IAB Tech Lab repository, contributed by Index Exchange under AGPL-3.0 while the specification carries a Creative Commons Attribution licence, listens for gRPC on port 50051, MCP on 50052, health probes on 8080 and a browser testing interface on 8081. Its supported intents cover segment activation, deal activation and suppression, deal floor and margin adjustment, bid shading and metrics addition.

From container project to finalised specification

The IAB Tech Lab launched its Containerization Project on June 4, 2025, with chief executive Anthony Katsur saying growth had pushed the framework programmatic is built on to its limits. Two months later, Zillow became the first brand advertiser to pilot containerised bidding with Chalice and Index Exchange, surfacing page-level signals that ordinary bid requests do not carry.

The specification entered public comment on November 13, 2025, with a comment window running to January 15, 2026. Sources disagree by a day: the specification document is dated November 12, as is PPC Land's companion report, while the IAB Tech Lab press release and most trade coverage use November 13. The work came out of the Container Project Task Force inside the Programmatic Supply Chain Working Group, with named contributors from Index Exchange, Chalice, The Trade Desk and OpenX, led by Miguel Morales and Shailley Singh.

On February 26, 2026 the framework was folded into a wider programme when IAB Tech Lab named its agentic initiative AAMP, the Agentic Advertising Management Protocols, placing ARTF at the base as the high performance foundation layer. Version 1.0 was declared final on August 12, 2026. According to Morales, now titled Director of Agentic Advertising, the final release added an orchestration layer letting one ARTF node call others and merge their answers into a single mutation set, a Rust reference implementation, alignment with OpenRTB gRPC, and clarified health endpoints. Work on version 2.0 is open for proposals.

Why buyers and sellers care

The commercial argument is about where computation happens. Demand-side platforms throttle the requests they accept because evaluating all of them on public cloud is unaffordable, so a buyer sees only a sampled slice of the market. Index Exchange chief executive Andrew Casale made the case in February 2026 for moving decisioning upstream to remove what he called the cloud tax.

Deployments followed. Bedrock Platform ran a full demand-side platform bidder inside an exchange on April 21, 2026, on Index Cloud infrastructure aligned with ARTF, with each container cryptographically signed by the partner and verified by the exchange, and an execution window Index Exchange puts under five milliseconds. PubMatic opened its auction to partner models through Decision Fabric on June 1, 2026. Magnite set out two deployment routes for outside models on August 6, 2026, containerisation among them. IAB Tech Lab states the framework cuts latency by 90%, a figure the organisation reports rather than an independent measurement.

Where the framework stops

ARTF validates the call, not the auction. Nothing in version 1.0 checks that a bid request remains valid OpenRTB after mutations are applied, so a remote procedure call can succeed while the resulting auction is incoherent. Vocabulary is inconsistent between the specification document, whose examples use names such as adjustDeals, and the protocol buffer definitions and reference implementation, which use constants such as ADJUST_DEAL_FLOOR.

The name is itself unsettled. The organisation has published the same standard as both Agentic RTB Framework and Agentic Real Time Framework, the licence section of the final document still uses the former, and its own August 2026 announcement dates the public comment release to February 2026 rather than November 2025.

A structural criticism concerns observability. Containerisation restores per-impression computation, but decisioning now happens inside a signed container that neither counterparty inspects by design, so the guarantee becomes the signature rather than the record. Rejection reporting is optional. Adoption costs engineering: containers, protocol buffer toolchains and Kubernetes are not standard equipment at most advertising vendors.

Adjacent terms

AdCP, the Ad Context Protocol, launched in October 2025 from a separate coalition and defines how agents describe campaign intent to each other. It sits above the auction; ARTF sits inside it. AAMP is the IAB Tech Lab umbrella containing ARTF, the Agent Registry and the buyer and seller agent SDKs. MCP is a transport ARTF may use, not a competitor to it. Index Cloud and Decision Fabric are vendor implementations aligned with ARTF, not the standard itself.

Recent developments

The finalised release arrived alongside a boundary dispute. On August 20, 2026 PPC Land reported that AAMP and AdCP overlap on thirteen functions, a count attributed to Singh, who rejected the description of AAMP as an impression-layer protocol. A second analysis of the same rebuttal noted that roughly ten months separate the two launches, too short a period for either to have built a settling installed base.

Timeline

  • June 4, 2025: IAB Tech Lab launches its Containerization Project
  • August 6, 2025: Zillow pilots containerised real-time bidding with Chalice and Index Exchange
  • November 12, 2025: Agentic Real Time Framework version 1.0 document is dated
  • November 13, 2025: IAB Tech Lab announces ARTF v1.0 for public comment, backed by Netflix, Paramount, The Trade Desk and Yahoo among others
  • January 15, 2026: Public comment period closes
  • February 26, 2026: ARTF is placed at the foundation layer of AAMP
  • April 21, 2026: Bedrock Platform runs a containerised bidder inside Index Exchange
  • June 1, 2026: PubMatic launches Decision Fabric for partner models
  • August 6, 2026: Magnite publishes its two routes for partner models inside its auction
  • August 12, 2026: ARTF version 1.0 is declared final, with an orchestration layer and a Rust reference implementation; version 2.0 work begins

Summary

Who. The IAB Tech Lab Container Project Task Force, inside the Programmatic Supply Chain Working Group, wrote the specification. Named contributors came from Index Exchange, Chalice, The Trade Desk and OpenX, with Miguel Morales and Shailley Singh leading for the organisation. Amazon Ads, Netflix, Yahoo, Paramount, Optable, HUMAN Security, Magnite, PubMatic, WPP Media and Basis Technologies also participated.

What. A specification for deploying partner logic as isolated containers inside a host advertising platform, communicating over gRPC or MCP, and proposing atomic changes to OpenRTB payloads through a patching protocol in which every change declares its intent and the host decides whether to apply it.

When. The Containerization Project opened in June 2025. Version 1.0 went to public comment in November 2025, closed comment in January 2026, joined the AAMP umbrella in February 2026 and was declared final on August 12, 2026. Version 2.0 is in development.

Where. Inside the infrastructure of supply-side platforms, exchanges and demand-side platforms, in the host's own data centres rather than on public cloud, across display, video, connected television and out-of-home inventory.

Why. Every external call in a real-time auction spends milliseconds that the auction does not have, and every bespoke partner integration costs engineering time that small vendors cannot fund. Moving the logic into the host removes both costs at once. The trade-off is that decisioning moves inside a sealed container that neither party inspects, which buys speed at the price of visibility.