What Happens to the Specialists?
The most serious objection to RACE Programming is not about speed or cost. It is about depth.
It goes like this. Our industry spent two decades specializing for a reason. A front-end engineer, a back-end engineer, a data engineer, a security specialist, a usability expert: each is a decision-maker in a domain that took years to master. RACE Programming runs on a small, senior, AI-augmented team with a deliberately limited set of roles. So the objection lands hard. You cannot collapse a security expert, a data expert, and a usability expert into two generalists and expect them to make the decisions those specialists make, even with AI helping. That knowledge does not compress.
I take this objection seriously, because it is half right.
What the objection gets right
Deep expertise is real, it took decades to build, and it does not disappear. RACE Programming does not claim it does. If anyone tells you a framework removes the need for a security specialist or a principal data engineer, walk away.
What it misses: the boundary moves
The objection assumes the boundary between a generalist decision and a specialist decision is fixed. It is not. It moves.
What AI tools change is not whether expertise matters. It is how many decisions require the expert to be in the room. In our experience, a senior generalist paired with capable AI tooling now handles roughly 95 percent of the decisions that used to require pulling in a specialist. Not because the generalist became a security architect overnight, but because the tools encode enough of the specialist’s reasoning to get the routine, well-trodden decisions right. The security review of a standard auth flow, the performance profile of a common query pattern, the accessibility pass on a familiar component: these no longer need a dedicated expert for every instance.
What stays with the deep expert
Two things, and they are the two highest-value uses of that expertise all along.
First, the rare and genuinely hard case. The novel threat model. The data architecture with no precedent. The usability problem no pattern library solves. These still need the person who spent a decade earning the judgment. A generalist plus AI does not know what it does not know, and on those rare hard cases that gap is exactly where things break. The expert is there for that.
Second, defining the standard. Your top experts set the bar that the generalists and the AI then execute against. The security specialist writes the threat-modeling standard. The staff data engineer defines the schema conventions. The design lead owns the usability principles. Once the standard exists, the generalist-plus-AI team holds the line on the majority of the work without the expert adjudicating every case.
What this does to the demand for experts
Push that 95 percent through and the demand for specialists changes shape. It does not vanish. It concentrates, and it moves up. The routine specialist work is absorbed by generalists and AI, so what is left is the top of the curve: the person whose threat model is genuinely novel, whose data architecture has no precedent, whose standard the rest of the organization executes against. You need fewer experts, and the ones you need are the very best.
That residual need becomes a dedicated role mainly at scale. Inside a single small team, the rare hard case and the occasional standard-setting are intermittent work, not a full seat. But when one organization runs tens or hundreds of RACE teams, a single top-tier security architect can set the standard a hundred generalist-plus-AI teams hold the line on, and be pulled in for the handful of genuinely hard cases across all of them. A three-person team rarely needs a full-time principal security engineer. A hundred RACE teams in one company do, and only one or two of them.
Moving the boundary, not removing the expert
So the RACE Programming answer to the specialist objection is not that we replaced the specialists. It is that we moved the boundary. The specialist is no longer consumed by routine decisions in their domain. They are freed for the rare hard work and for setting the standards, which is where their decades actually pay off. The generalist team, armed with AI and those standards, covers the rest.
This is not all-or-nothing, and anyone selling it as all-or-nothing is overselling. Expertise must still exist. It must still be around. But it is not needed for the majority of the job anymore, and pretending otherwise, in either direction, misreads what AI actually did to the work.
The interesting question is no longer whether two people plus AI can replace ten specialists. That framing is a trap in both directions. The real question is where the boundary sits today, and which decisions have crossed from “needs the expert” to “the generalist and the tools have this.” That boundary moves every quarter. RACE Programming is built for teams that want to operate right at it.
Written by Pavel Khodalev, author of RACE Programming and CTO of First Line Software. Follow new essays via RSS.