The phrase "agentic media buying" now appears across vendor materials, conference programmes, and analyst reports. It describes a wide range of things. In some contexts it means an automated system executing pre-approved rules. In others it means a language model conducting open-ended negotiations on behalf of a buyer. The gap between those two descriptions is large enough that "agentic" has become nearly meaningless as a descriptor without further qualification.

Concourse Bench v1 (concourse.agency), published by Alkimi, provides a more specific reference point. It does not define agentic media buying for the industry. It describes, in recorded and documented detail, what one category of AI buying system actually does in a structured media-buying workflow.

What did the AI buyers in v1 actually do?

Each model-driven buyer in v1 received a fictional media-buying brief, a fixed budget, and access to three sellers running fixed software policies. It negotiated across multiple rounds with each seller to secure packages that could be combined into a portfolio within the registered completion rules. The system was evaluated as a whole: the model together with the buyer harness, the software that supplied the brief, checked every action, and recorded the results.

The actions the system performed were specific: read the brief, initiate contact with sellers, make opening offers, respond to counter-offers, evaluate budget constraints across the portfolio as negotiations progressed, walk away from sellers when terms did not work, and commit to a final portfolio. Fifty-five of 96 registered attempts resulted in a completed portfolio. Two models completed every attempt they ran. Two completed none.

This is a defined set of actions within a constrained environment. It is a more specific thing than "agentic media buying," which is why the specificity matters.

What distinguishes a model-driven buyer from an automated rule engine?

The distinction that matters most for practitioners is whether the system makes decisions in context or executes pre-specified instructions. An automated rule engine executes a decision tree: if the price is above X, decline; if viewability is below Y, issue a counter-offer at Z. A model-driven buyer in the Concourse framework does something different: it reads the brief, parses the sellers' responses, evaluates options against constraints it was not explicitly programmed to enumerate, and produces decisions that vary by context.

V1 records actions rather than the models' private reasoning, so the mechanisms behind individual decisions are not published. But the variation in completion rates between models, and between buying situations, indicates that different models respond differently to the same conditions. A rule engine with the same rules would produce the same outcome regardless of model. That responsiveness to context is the functional distinction.

What does earned autonomy mean for a media-buying deployment?

The concept of earned autonomy describes a progression from observation through to bounded automatic action, where each stage of autonomy is earned through demonstrated performance rather than assumed at deployment. It is a recognition that full autonomy in buying is not the right starting point for most deployments, and that the evidence needed to justify increasing autonomy must be gathered incrementally.

In practice, a sensible deployment path does not begin with the AI running the full buying cycle unsupervised. A more defensible starting point is the AI generating recommended portfolios that a human reviews and approves, before moving to the AI committing to pre-approved parameters autonomously within defined limits. The data on completion rates and contractual quality from v1 is precisely the kind of evidence that would inform that progression: a model that completes reliably and secures strong contractual terms has earned a different level of trust from a model that completes inconsistently or accepts weak viewability floors on 86% of its contracts.

What is the buyer harness, and why does it matter?

Concourse v1 evaluates model-harness combinations, not models alone. The buyer harness supplies the brief, checks every action against the registered rules, and records the outcomes. In a real deployment, the equivalent is the software layer that connects the AI to the trading environment, enforces the brief, checks that proposed contracts meet the buyer's criteria, and produces the record of what was committed and why.

The harness is not incidental to the result. The benchmark evaluates the system, not the model in isolation, for this reason. A model that performs well with one harness may perform differently with another. For practitioners building AI buying systems, the design of the harness layer is as consequential as the model selection.

What does v1 establish about the meaning of agentic buying?

V1 establishes that some AI models, in controlled simulation conditions, can complete the core actions of a media-buying workflow: read a brief, negotiate with multiple sellers, manage a budget constraint across a portfolio, and commit a completed set of deals. Two models did this reliably across every situation tested. Others could not.

What the benchmark does not establish: that any of these systems can be deployed in live markets without oversight, that completion in simulation predicts completion in production, or that the models that completed reliably also negotiated strong contractual terms. The viewability data, showing 95 of 110 contracts in completed buys with floors below 70%, indicates that completion and quality are separate dimensions that do not move together automatically.

"Agentic media buying" is a meaningful phrase when it refers to a model-driven system that takes consequential actions within a defined brief and buying environment, with a human retaining approval rights at defined decision points and with a record of what was committed and why. It is a less useful phrase when applied to the broader aspiration of full autonomy across the buying cycle, because the evidence base for evaluating that aspiration is still being built.

What comes next in the benchmark?

Concourse Bench is developed by Alkimi, which builds trading infrastructure for AI agents in media markets. V2 will test buying quality more strictly, with different tasks and rules. V2.1 will introduce a model-driven seller to run against the buyers. V2.2 will study buyer and seller models interacting.

The full data, methodology, recorded simulations, and definitions from v1 are available at concourse.agency. Practitioners evaluating AI buying systems can use the benchmark as a reference point for what a structured evaluation of completion and contractual quality looks like, and what the results indicate about which models are ready for deployment consideration under which conditions.