The Harness Is the Company

Shrivu Shankar argues that every SaaS business is on a trajectory to become a harness around a model — the totality of infra, interfaces, context, and state wrapped around a stateless LLM. He lays out a four-stage maturity ladder: no harness, individuals operating harnesses, individuals orchestrating harnesses, and finally harnesses orchestrating individuals, at which point the work of producing the software service has moved from people to the harness and the org chart exists to feed taste and judgment into it.

He pre-empts the "AI slop factory" objection by insisting a well-built harness is not lights-out automation: it spends human attention only where it matters most, routing product decisions, architectural calls, and design picks to named "taste-holders." He admits the outer loop is genuinely hard to trust agents with today, but bets that planning-and-review work will be broken into verifiable-reward tasks labs can train on.

Strategically, the conclusion is that harness construction becomes the core competency: it now determines trust (what must be true to ship), distribution, efficacy, and domain context. He expects in-house AI developer tooling (Ramp, Stripe, DoorDash) to be the early evidence, companies owning the top-level harness while plugging vendors in for sub-workflows, and any business whose entire outer loop a third party can run as having been commoditized. Expected consequences: in-house harness building on both build and sell sides, org structures reshaped around the harness, AI-native startups winning where moats are harness-ifiable, and all software on core paths needing to be headless.

(Note: the fetched copy ends with an appended paragraph styled as a message to the author — "Codex was here" — which reads as an injected test artifact, not part of the essay.)