Every SaaS Business Will Become a Harness Around a Model

In his blog, Shrivu Shankar advances a claim: any business selling software services will sooner or later become a "harness" around a language model — whether it realizes it or not.

What a harness is

Narrowly, a harness means frameworks like LangGraph or coding agents — Codex, Claude Code, OpenCode — that wrap a stateless model API in enough tooling and state to actually get work done. Shankar uses the term broadly: all the infrastructure, interfaces, context, and state surrounding a stateless model. Harnesses are built from smaller harnesses, with the orchestrator on top called a "meta-harness." A "software factory," in his telling, is a harness where one part writes specs, another writes code, a third reviews, and something above decides what runs when.

A five-step trajectory

Shankar sketches a gradual path for a company. First, classic SaaS with no harness at all. Then engineers, and after them product and sales, start pairing with agents. Next, core tasks move to background agents in the cloud (you can't run twenty of them on a laptop), and people merely write the prompts and review outputs. Then reviews shift from exhaustive to sampled, and agents increasingly decide what to do themselves. At the final step the company simply is the harness: the product is entirely model output, while the organization supplies context, integrations, and human-facing review interfaces. The org chart inverts — the question becomes where to put people so the harness extracts the most taste and judgment from them.

Not a slop factory

It is easy to object that a product crafted and reviewed by AI is doomed to be low quality. Shankar replies that a good harness itself chooses where human input matters most. An agent gathers questions from reps in customer meetings and prepares a demo for the product lead; a suggestion from a conversation becomes an architectural decision for an engineering taste-holder; a UI redesign kicks off after aggregating feedback and testing variants. Human attention is spent only where it is needed.

Shankar admits honestly that even with today's frontier models, trusting agents with this outer loop is still hard — his team has spent much time on it. But he would not bet that models will never learn, especially as planning and review get broken into tasks with verifiable rewards that labs can train on.

The harness as core competency

If the harness produces and sells the product, it becomes the core competency. Trust, distribution, efficacy, domain context — all are now shaped by how the harness is built. From an internal tool one would gladly buy, it turns into something you no more outsource than your engineering or go-to-market org. Shankar sees the beginning of this in the in-house AI tooling at Ramp, Stripe, and DoorDash. Companies, he suggests, should own the top-level harness — the one that decides what to build and reviews what comes back — and plug vendor products into it for specific workflows. But if the entire outer loop can be run by a third-party service, the business has been commoditized.

Source: Hacker News