Selling to the Buyers Who Don't Exist Yet

Flynn's "How to Sell to Agents" opens with Coase's Nobel-winning insight — firms exist because transaction costs make markets expensive — and argues that AI agents collapse those costs so dramatically that the logic of buying versus building flips entirely:

An agent can discover a service, check its price, and call it in a single HTTP roundtrip. No proposal. No demo. No comparison-shopping across ten browser tabs.

This is a good frame. The buy-vs-build arithmetic is the best part: a general-purpose agent spending $0.10–0.50 and 25 seconds reasoning through a sub-task versus a specialised endpoint returning the same answer for a penny in 200 milliseconds. That's not a marginal improvement. That's a category difference. The insight that you don't sell intelligence to agents — they have plenty — you sell access to things they can't compute on their own is the kind of line that ages well.

But then, almost in passing:

Early data from agent service catalogs paints a stark picture. In one sweep across 44 services, only 2 had fully working endpoints. 53% of direct service calls succeeded.

Two out of forty-four. That's a 4.5% fully-functional rate. For an ecosystem the article describes as if it's one architectural decision away from reshaping commerce.

The piece knows this is a problem and frames it as opportunity: be the reliable one and you win. Fair enough. But there's a difference between "the infrastructure isn't ready yet" and "the infrastructure is nearly ready and the smart move is to build for it now." The article's tone is firmly the latter. The data says the former.

The other thing. The article declares the attention economy irrelevant — agents don't browse, they query, no brand loyalty, no impulse purchasing — and then immediately concedes that a human still decides which tools the agent is allowed to use. That human absolutely responds to brand, reputation, word-of-mouth, and yes, marketing. The runtime is automated. The allowlist is not. Saying attention doesn't matter and then walking it back two paragraphs later isn't nuance. It's a thesis that couldn't survive contact with its own argument.

None of which means the direction is wrong. Machine-readable capability registries, per-request pricing, automatable onboarding — these are probably where B2B service architecture ends up. The Coase logic is sound. HTTP 402 finally finding its purpose after three decades in the spec is a good story because it might actually be true.

But "agents are already spending money" and "agents will eventually spend money under tightly constrained enterprise governance policies after several years of procurement process adaptation" are very different statements. The article sells the first. The reality, for now, is the second.

The future it describes is probably right. The present it implies is not.