Agentic Buying Doesn't Fail at Intelligence. It Fails Without a Harness

Buying is going agentic faster than verification is

The IAB's 2026 Outlook Study puts numbers on what every trading desk already feels: 96% of buyers are aware of agentic AI for ad buying, and 66% are increasing their focus on it for buying and campaign execution this year. At the same time, 40% of those buyers name understanding agentic AI as an investment challenge.

Video is moving even faster. In the IAB's 2026 Digital Video Ad Spend & Strategy Report, two in three digital video buyers are already live (21%), testing (20%), or planning to use (25%) agentic AI for video campaigns, with another 28% actively investigating.

So execution is being automated at speed. What is not being automated at the same speed is verification: the step that decides whether what the agent produced will actually deliver, play, and track. That gap is where agentic buying breaks.

What coding agents already proved about autonomy

Software engineering went through this exact transition first, and the result was unambiguous: coding agents did not become useful because models got smarter in isolation. They became useful because they run inside a harness: a compiler, a type checker, a linter, a test suite, CI. Every action the agent takes gets deterministic, machine-readable feedback, and the agent iterates against that feedback until the work passes.

Take the harness away and an LLM writing code is just producing plausible-looking text. Nobody ships that.

A media-buying agent assembling a deal and trafficking a video tag is in the same position, except worse: there is no compiler for a VAST tag. XML that is well-formed can still reference a deprecated VPAID creative, omit a required Duration, carry an insecure tracker, or bury the media file four wrappers deep. The agent needs the same loop a coding agent gets (act, validate, repair, re-validate), and that loop requires a deterministic validator it can call as a tool.

The standards bodies are building the protocols. Protocols are not verification

The industry's standards work confirms the direction. In January 2026, IAB Tech Lab released AAMP (Agentic Advertising Management Protocols) v1.0, an umbrella framework for agent-driven advertising that deliberately builds on existing standards: OpenRTB, AdCOM, OpenDirect, and VAST, wired together with modern agent protocols like the Model Context Protocol and Agent2Agent. Tech Lab's own framing is blunt: without speed, structure, security, and trust, there's no agentic advertising future worth having.

A few months earlier, in October 2025, a consortium including Yahoo, PubMatic, Optable, Scope3, Swivel, and Triton Digital launched the Ad Context Protocol (AdCP), an open standard built on MCP that lets buyer and seller agents exchange context and orchestrate buys across platforms.

Both efforts solve the communication problem: how agents discover inventory, negotiate, and transact. Neither one verifies that the payload changing hands is correct. When a human trafficker is removed from the loop and both sides of a transaction are machines, nobody is left to eyeball the tag. Protocol adoption makes payload validation more urgent, not less. It is the missing rung between agents that can transact and agents that can be trusted to.

Programmatic already showed what automation without verification costs

We do not have to speculate about what happens when execution is automated and verification is not, because programmatic already ran the experiment. The ANA's Programmatic Transparency Benchmark found $26.8 billion in wasted global programmatic spend in its Q2 2025 report, up 34% from the roughly $20 billion its landmark 2023 study identified. The same research found the average campaign running across 44,000 websites when a few hundred would reach most of the audience.

Programmatic automated the buying loop and left quality checking to quarterly reports and post-campaign audits. Agentic buying shortens the decision loop from hours to seconds. Run that loop without inline verification and the waste does not shrink. It compounds at machine speed.

What the harness layer has to provide

• Deterministic results: the same input produces the same findings every run, with stable rule IDs an agent can branch on.
• Machine-readable output: structured JSON and exit codes that can gate a pipeline, not prose a human has to interpret.
• Spec grounding: every rule traceable to a section of the IAB VAST, OMID, or SIMID specs, so a failure is a standards violation, not a vendor opinion.
• Availability everywhere agents run: as a CLI for pipelines, as native libraries inside ad servers, as a CI action, and as an MCP tool agents can call directly.
• Wrapper-chain resolution, so what gets validated is what the player will actually receive, not just the first hop.
• Bounded auto-repair: deterministic, low-risk fixes only, always followed by re-validation.
• One shared gate for humans and agents, so QA standards do not drift between the two.

Where vastlint sits in that stack

Video is where the payload is most fragile and most expensive. The IAB projects U.S. digital video ad spend to surpass $80 billion in 2026, with CTV alone at $29.3 billion, and VAST is the delivery contract underneath nearly all of it.

vastlint is the lint layer for that contract: 129 deterministic rules mapped to the IAB VAST, OMID, and SIMID specs, with stable rule IDs, NDJSON output, wrapper unwrapping, and a deliberately conservative fix mode. It runs as a CLI, as Go and Rust libraries, as a GitHub Action, and as a hosted MCP server: the same gate whether the caller is an ops analyst, a CI job, or a buying agent.

To be clear about scope: payload validation is one slice of the harness, not the whole thing. Agentic buying will also need gates for brand safety, supply path, and pacing. But the creative delivery payload is the most compiler-like piece of the problem: fully specified by IAB standards, deterministically checkable. It is the slice you can put in front of an agent today.