- Published on
Every SaaS business will become a harness around a model
- Authors
- Name
- DeepHorus
A sweeping prediction is emerging from the AI software community: every SaaS business will eventually become a harness around a model, whether or not its leaders have realized it yet. The argument holds that as large language models grow capable, the value of software companies will shift from the code they ship to the infrastructure, context, and review systems they wrap around stateless models. In this framing, the model itself becomes a commodity, and the durable business lives in everything surrounding it. It is a claim that reframes both product strategy and organizational design for the software-as-a-service industry.
The essay introduces the term "harness" broadly to mean all the infrastructure, interfaces, context, and state surrounding a stateless LLM, not merely developer frameworks such as LangGraph or coding agents like Codex, Claude Code, and OpenCode. A harness can be composed of task-specific sub-harnesses coordinated by a "meta-harness," and a "software factory" is described as a harness whose parts are smaller harnesses: one writing specs, one writing code, one reviewing, with a layer on top deciding what runs when. This broader definition is what allows the author to map an entire company onto the concept. It also suggests that what developers already do with agent tooling is a preview of what whole organizations will do.
The proposed trajectory moves through four stages. First, companies sell software services built the traditional SaaS way, with no harness. Second, engineers pair with agents, and other functions like product and sales follow, meaning individuals operate harnesses. Third, core tasks shift to background agents running in the cloud β a laptop cannot run twenty of them β and employees trigger and review them, meaning individuals orchestrate harnesses. Fourth, proactive background agents take over, continuous review gives way to sampled review, and harnesses orchestrate individuals.
Following this path, the company itself becomes the harness: core tasks are performed by agents, the product is largely model output, and the business supplies context, integrations, and human-facing review interfaces. The relationship between software and org structure inverts, with the org chart reorganized around where people can give the harness the most taste and judgment. Domain knowledge, tooling, permissions, review loops, and context all consolidate into a single business harness. Humans do not disappear; they become components of it.
Critics may call this an AI slop factory, assuming products crafted and reviewed by AI are inherently low quality. The author counters that this assumes a lights-out operation, and that the core mitigation is having the harness decide where human input matters most. Examples include an agent fanning out questions to sales reps during customer meetings before synthesizing a demo for the product lead, or surfacing a feature request as an architectural decision for an engineering taste-holder. A good harness maximizes customer value while spending human attention only where it is needed.
Skepticism remains understandable. The author concedes that even with today's frontier models and a well-crafted harness, trusting agents to handle the outer loop of planning and review is difficult, a challenge his team has spent considerable time attempting. The counter-argument is strategic rather than technical: it is not worth betting that models will never manage this work. As planning-and-review tasks break into units with verifiable rewards, they become trainable for frontier labs.
For SaaS operators, the near-term question is where to place human judgment rather than how to eliminate people. Companies that design explicit review loops and accumulate proprietary context may find the harness becoming their moat. Those that treat agents as simple autocomplete risk commoditizing their own differentiation. The prediction implies that competitive advantage will accrue to whoever builds the most effective wrapper around increasingly capable models.
