The end of interfaces, not the end of software

Interfaces, yes. Software, no. The distinction looks thin on paper and it decides the product strategy of every vendor for the next five years.

The number that started the debate comes from Supabase, and it is worth quoting in full rather than in the shortcut that travelled. In July 2026 its co-founder Paul Copplestone told the Y Combinator podcast: “we can measure 60% of databases get launched by an agent, but more likely it’s like 90%, and it’s in the millions every month.” The platform went from 6.5 million developers at the end of 2025 to 10 million by July 2026, and its single largest contributor is not a company but a coding agent.

A product whose users mostly never see the screen

Take the measure of what that means for a product team. You design journeys, you polish a dashboard, you test button labels, and most of your new users will never open any of those screens. They will read your documentation, call your API and leave.

Supabase is a favourable case, being an infrastructure product aimed at developers. But the direction is general and already visible at vendors with nothing technical about them: Salesforce has exposed its entire platform through APIs so an agent can read data, trigger workflows and update records without ever touching the interface, and firms report folding five or six separate subscriptions into a single conversational surface.

As for how many agents are actually out there, nobody knows. Forbes concluded in June 2026 that their number sat somewhere between 10% and 100% of the human population, which is a range wide enough to signal mainly that we are not yet measuring the right thing.

What dies is not the software, it is the head

The phrase circulating among investors is “headless” software. The system of record stays exactly where it was, because the data still has to live somewhere, the rules still have to apply somewhere and the history still has to be kept. What detaches is the single screen you reached it through, replaced by as many interfaces as there are ways to phrase a request.

That shift has a brutal consequence for design. An agent does not care about visible features, it cares about reachable capabilities: ten well-designed API endpoints are more useful to it than two hundred dashboard screens, and that sentence inverts roughly everything product management has learned in fifteen years.

It also explains why some visually rich products will turn out to be surprisingly poor in agentic use. A feature that exists only as a click, a state that can only be read visually, an action that requires three successive screens: none of it is usable by a system that does not look.

The word everyone avoids: identity

Here is the real open problem. If your users become machines, under whose name do they act, and who granted that name?

The question sounds administrative and it is structural. An agent creating a database on Supabase does so with somebody’s credentials. An agent writing into your ATS commits the company that opened the access. Yet most current systems can only model two situations, the human user and the shared service account, and neither describes what is actually happening. That is precisely the problem NIST named in February 2026, observing that enterprise agents inherit either all of their user’s permissions or an identity nobody can audit, and it is what makes a successful injection so hard to detect.

Our conviction is that this will become the main selection criterion for professional software within two years, well ahead of feature depth. A vendor able to say precisely which identity did what, with which permissions and on whose request, will be worth more than one with a prettier interface.

What the screen keeps, and nobody mentions

It would be dishonest to predict interfaces disappearing entirely, because three uses resist and none of them is marginal.

The first is exploration. A conversation is excellent for getting what you know how to ask for and very poor at surfacing what you do not know exists. Nobody ever found an anomaly in their numbers by asking a question: you find it by looking at a table and noticing something is off. Until an agent knows what will surprise you, the screen remains the only place surprise can happen.

The second is verification before commitment, and it is the one that concerns us most directly. The queue where a person reviews what an agent is about to send is an interface, an entirely conventional one, and it cannot be replaced by a conversation: approving requires seeing. It is in fact the only screen we consider non-negotiable in Balt.

The third is trust over time. A system whose state you never see eventually makes people uneasy even when it works, and unease costs more than the screen it would have taken to prevent.

So the honest way to frame the transition is not that interfaces disappear, but that they stop being the mandatory point of passage. They become one mode of access among several, reserved for what they do best, which is already a considerable change for anyone who built their product around them.

What it changes concretely for a staffing firm

Three things, and none of them requires waiting for the market to finish moving.

Your ATS stops being judged on its ergonomics, which is good news for everybody. The question becomes whether its capabilities are reachable other than by a click, because a tool that does not open up becomes an island where your data is stranded while the rest of your systems talk to each other. It is also what puts reversibility back at the centre of choosing a vendor.

Per-user pricing becomes an anomaly your suppliers will defend less and less well. Charging a machine for a seat makes no sense, and the model is already sliding towards what is consumed or accomplished, exactly as per-seat pricing turns against an agent vendor.

Finally, your own internal screens lose relative importance. The place where work gets requested becomes the inbox, and that shift applies to your consultants as much as to your clients.

Our bet, and what it costs us

We built Balt assuming the interface is not the product, which is easy to say and expensive to live with. In practice it means we invest little in screens and a great deal in what sits behind them: the quality of the context, the precision of the permissions, the trace of what was done.

The commercial cost is immediate and we see it in every demonstration. A product whose substance is invisible sells less well than an impressive dashboard, and it takes longer to convince somebody that the absence of a screen is a decision rather than a gap.

We still think it is the right bet, for a simple reason. When intelligence tends towards zero and scarcity moves to context and the right to act, the software that keeps its value is not the one that displays best, it is the one people agree to trust with something. An interface can be rebuilt in six months. A permission is earned once and lost for good.