An attendee at NYC's AI Agents Conference argues most companies are betting on defensibility strategies that won't survive. The analysis suggests prompt engineering and domain expertise won't protect against commoditization when engineering labor approaches free.
I've been to dozens of these conferences. I know the pitch: we're building the picks and shovels for the AI gold rush. We bundle expensive engineering and domain expertise into tools that solve real problems. Except this time, engineering is becoming free and domain expertise is text that can be copied.
The conference floor was packed with observability platforms, governance tools, supervisor agents, and data substrates. Companies selling solutions to problems that emerged this year when agents hit production. Everyone was implicitly betting on a new moat to replace the old SaaS model.
The old pitch was simple: we bundle R&D you can't afford into tools you can. But in a world where AI can vibe-code much of what these booths were selling in days or weeks, that moat evaporates. The expensive part was always the engineering. Now engineering is approaching free.
According to a detailed Reddit post from an attendee, one VC speaker said his key metric for evaluating AI-native startups is ARR per engineer—and that number should be climbing. Translation: fewer humans generating more revenue, because the AI does the work.
The most popular bet at the conference was on encoded domain expertise. Hire lawyers to build legal AI, doctors to build medical AI, engineers to build developer tools. The expertise is the moat, the thinking goes.
I'm skeptical. Prompt architecture is text. It's portable. The expertise underneath is often abundant—there are over a million lawyers in the USA alone. The righteous destiny for this category is open marketplaces of prompt architecture, not trade secrets.
Another popular approach: data substrate companies that wire up databases, pipelines, Slack threads, and tickets into context graphs agents can reason over. Connect a database, watch an agent crawl the schema and produce a chatbot interface. It feels magical.
But strip the magic away and most of these are prompt architectures on top of LLMs plus data ingestion. You're defending yourself against open-source alternatives and your customer's internal team.
