Every SaaS endeavor volition rotate into a harness about a model, whether or not they’ve realized it yet.
Narrowly, folks volition frequently affiliate a “harness” alongside frameworks akin LangGraph or coding agents akin Codex, Claude Code, or OpenCode, which cover a homeless example API in adequate tooling and province that you can really get activity done.
I’m using the term “harness” broadly to average all the infra, interfaces, context, and province that encircle a homeless LLM. A harness can be composed of adapted or task-specific sub-harnesses (with the orchestrator being called a “meta-harness”). A “software factory” is a harness whose parts are smaller harnesses — one that writes specs, one that writes code, one that reviews — affirmative item on top deciding what runs when.
If you obtain this broader definition (or replaceAll “harness” alongside any expression you’d prefer), I doubtful for many application assistance businesses you’ll see this trajectory:
They market application services built the traditional SaaS-y way.
No harness.
They market application services, but the engineers brace alongside agents to get the activity done. Increasingly another functions akin merchandise and revenue brace alongside agents for productivity.
Individuals run harnesses.
They market application services alongside many center tasks moving to backdrop agents operating in the haze (a laptop can’t run twenty of them). Engineering, product, and revenue all trigger these agents (i.e. compose the prompts) and assessment the outputs.
Individuals orchestrate harnesses.
They market application services alongside many center tasks moving to proactive backdrop agents, alongside engineering, product, and revenue moving to reviewing delegate outputs. Consistent reviews move to sampled reviews. Increasingly the agents decide what to do proactively, alternatively of humans designing the activity up front.
Harnesses orchestrate individuals.
Following this trajectory, you’ve really turned your business into a harness. The activity to create the application assistance has moved from group to the harness.
Core tasks are done by agents and the “product” is now entirely example output, alongside the business supplying the context, integrations, and human-facing assessment interfaces.
The association between application and org construction has inverted as the org diagram becomes a inquiry of anywhere to put group so the harness gets the most sensation and judgement out of them. Humans are part of the harness.
The company’s domain knowledge, tooling, permissions, assessment loops, context, etc. all rotate into the endeavor harness.
It’ll be tempting to think that products that are chiefly crafted and reviewed by AI volition be inherently low norm and that construction through a harness alternatively of through group volition create a slop factory at scale. I think that condemnation assumes a business run this way is lights-out.
The center mitigation is really having the harness choice anywhere individual inputs matter most. For example:
A merchandise decision could arrive from an delegate fanning out questions to reps in client meetings and afterward synthesizing a merchandise demo for the merchandise guide to review
A characteristic recommendation pulled from a client gathering is turned into a important architectural decision presented to an engineering taste-holder by the agent.
A UI redesign kicks off following aggregating feedback, and following evaluation a few variants, the top options are presented to a scheme taste-holder.
A fine harness maximizes value to the client during spending individual notice — employees’ and customers’ — lone anywhere it’s needed.
If you are motionless skeptical, that’s fair. Even alongside today’s frontier models and a well-crafted harness, it’s fairly difficult to rely agents to grip the outer loop akin this (we’ve spent a lot of period trying). I fair don’t think it’s value betting that models won’t be capable to eventually do this, particularly as additional of this planning-and-review activity gets damaged into tasks alongside verifiable rewards that labs can train on.
To examine additional on what an institution of “taste-holders” could appearance like, see The Transposed Organization. This article builds fairly a bit on the ideas laid out in that article.
Depending on the service, the differentiation a business has frequently comes from things akin trust, distribution, efficacy, and domain context. In a proactive backdrop delegate world, your capability to build the harness — how it learns, what it watches for, how it interfaces alongside individual taste-holders, what systems it integrates alongside — volition additional and additional be how you keep that differentiation.
The harness is now what shapes what must be true for activity to container (~trust), how accelerated and how the products district (~distribution), the speed and environment of feedback loops (~efficacy), and how organizational cognition is ingested and maintained (~domain context).
Harnesses go from inner tooling you’d happily buy to item you’d no additional outsource than your product-eng org or your GTM team. This is distinctly distinct from the pre-AI world, anywhere outputs were mostly bounded by the humans using the application to get things done.
I think in-house AI developer tools (like the ones we’ve seen from Ramp, Stripe, DoorDash, etc.) are the commencement of this.
For AI-pilled companies, waiting for an SDLC tool vendor to add an integration, assistance a certain interface, or attain a flat of cost-efficacy increasingly bottlenecks their capability to build and keep their product. This is particularly true in the abbreviated term for third-party tools that can’t yet run an enterprise’s entire application factory for it (often since the tech stack is too bespoke, governance too restrictive, crucial characteristic assistance too slow, or because of a preferred disbursal model).
I don’t anticipate everything to be built and maintained in-house. Rather, companies should own the top-level harness — the one that decides what to build and reviews what comes rear — and plug vendor products into it for particular workflows. Eventually a third gathering volition get beautiful fine at an enterprise-level “spec to tested drag request” and at that item a business can toggle out that part of the application iteration alongside that merchandise during motionless maintaining the agent(s) that compose the input spec and grip the next steps from drag petition output.
If somehow the complete external iteration can be done by a third-party harness (i.e. operating the complete endeavor via proactive backdrop agents as a service), afterward I’d contend the endeavor has now been commoditized.
If this is the correct mental model, you should anticipate to see:
An different amount of in-house harness construction on the two the build flank and the market side
Org structures and idiosyncratic roles being reshaped about their location in the endeavor harness
AI-native startups beating incumbents in domains anywhere the “moat” can be effortlessly harness-ified
All application a application business uses (on or tied to the center build or market paths) needing to be headless so the external harness can run it
