The Harness, the Horse, or the Hay
A framework for understanding the AI stack: you're either the LLM (Horse), the application layer (Harness), or infrastructure (Hay)—and most founders are running the wrong playbook for their category, which is what gets them killed.
Read Original Summary used for search
TLDR
• The AI application layer consolidates to ONE app per job function, not "fewer apps"—75% of employee-facing software companies will disappear
• Three viable models exist: Horse (LLM), Harness (application owning ontology + loop for one rider's entire job), or Hay (infrastructure serving builders)
• The fatal mistake isn't picking the wrong category—it's running the other category's playbook (harness makers expanding cautiously, hay companies selling transformation to execs)
• "No partial harness" rule: build one job-to-be-done incompletely and you're a feature, not a competitor—customers hand over adjacent systems within 60 days if you nail the first job
• Harness playbook: own the rider's entire week, expand by context not category, price the transformation. Hay playbook: sell instruments not judgment, pool what individuals can't, be undeniable at one thing first before bundling
In Detail
The AI application market is consolidating to one app per job function, forcing every company to choose between three positions: the Horse (LLM model), the Harness (application layer), or the Hay (infrastructure). Model providers won't win the application layer despite their advantages—customers won't wed themselves to one provider, the specialization required doesn't scale economically, and no serious organization will hand its differentiated intelligence back to a platform serving competitors. The real battle is between harness and hay companies running each other's playbooks.
A harness must do four things: carry the ontology (the meaning between data, not just schema), own the interaction model for that role, run an eval/RL platform where graded outcomes change behavior without engineering, and be singular—one master agent per rider. There is no partial harness. Build three jobs while a competitor builds fifteen and you're not smaller, you're a feature they'll ship in a sprint. The sequencing rule: build jobs that widen context not invoices, stay adjacent in time not category, prefer gradable work, and ensure each new job makes previous jobs better. Customers volunteer to hand over adjacent systems within 60 days, asking "why don't you just build that?" The harness sells transformation to executives and wins by capturing imagination before competitors arrive.
Hay is everything the horse and harness need at scale: ingestion, substrate, memory, serving, execution, evals, guardrails, identity. The playbook inverts completely. Your buyer is a builder evaluating you alone at night in fifteen minutes. Sell the instrument that produces judgment, never the judgment itself—own the trace store and eval runner, not the definition of correct. Pool something no single customer can pool (cross-provider pricing, failure signatures, attack patterns). Price the agent not the seat. Land while stacks are molten, because in 18-36 months harness winners will have ossified and switching stops. Be undeniable at one thing before bundling, because gravity-driven consolidation happens by acquisition, and mediocrity at six things gets you absorbed. The downside case in hay is a sale; in a partial harness it's zero.