ARTF: the idea is right, the contract works against it

IAB Tech Lab’s Agentic Real-Time Framework (ARTF) is one of the most interesting ideas in programmatic right now. Not because of “AI in the auction”, but because it turns the auction into a platform.

The idea

Today every new piece of auction logic (floors, deal curation, segments, bid shading, fraud checks) is a custom integration. ARTF replaces that with one contract for agents from any vendor:

• an agent ships as a standard container, with a manifest declaring its intents and resources;
• it never writes to the bid request, it proposes typed mutations: intent, operation, path, value;
• the host’s orchestrator decides what to apply and keeps the data, the SLA and the deadline.

This is why gRPC matters here: the contract doesn’t care where the agent lives. A sidecar next to the auction, a separate service in the cluster, or a remote endpoint when the latency budget allows: same interface, same validation. A vendor packages its logic once and runs it on any compliant exchange, without seeing more data than it’s entitled to.

Where the contract still hurts

1. Protobuf in a JSON world. ARTF mandates protobuf with its own copy of the OpenRTB 2.x schema, while most of the industry still runs OpenRTB as JSON. Every call starts with mapping one model into the other, on a millisecond path.
2. Extensions eat portability. Data that isn’t in the bid request goes into protobuf extensions with field numbers from 500 up, claimed in blocks per organisation. An agent that depends on one host’s numbers is an integration again.
3. No per-DSP stage. The protocol defines two lifecycle points so far: the publisher’s bid request and the DSP’s bid response. A mutation to the request applies to every partner, while useful optimisations are often partner-specific.
4. One impression, one floor. An impression can offer banner, video and native at once, but it carries a single floor, and the floor mutation takes one number. A model that prices formats differently has nowhere to put its answer.

None of this kills the idea. It’s the price of a v1.

The part I find most promising

Once every agent goes through one orchestrator, with declared intents and measured acceptance of its mutations, the auction gets a universal control point. Switch agents on and off, schedule them, tune parameters, compare, roll back: for any agent, from any vendor, without touching the hot path.

That’s a natural base for automated control. First rules, then models, and eventually an AI agent connected over MCP. ARTF already supports MCP, but as another way to call an agent. I’d put it on the orchestrator: an LLM is too slow to sit inside a millisecond auction, but it’s a good fit for managing the system that runs one.

How do you see it: will exchanges open their auctions to third-party agents, or keep ARTF as an internal plug-in system?