Boris is naming the shift my own business has been built around for the last few months The old way to scale an agency: hire a junior, train them on your last 10 projects, hope they retain enough that the 11th project doesn’t start from scratch Repeat with a senior to manage the juniors Repeat with a VP to manage the seniors Every layer adds overhead, every layer dilutes the knowledge that lived in the founder’s head. By the time the agency is 30 people deep, the original quality is gone and the founder is doing nothing they trained for The new way: you accumulate the data from every build into a structured library, then you write loops that run against it. Agents looking at past patterns, past objections, past architectural decisions, past failure modes. New client request comes in, the loops propose what to ship before anyone on the team has read the brief. Not from theory, from 50 builds the loops have already processed. The headcount math breaks. Hiring scales linearly with overhead. Loops scale with the size of the data library, which compounds with every build. What this looks like in practice for me right now: Every client kickoff feeds into a structured project log Every architecture decision is tagged and described Every issue we hit in production gets logged with the fix Every prompt that worked on a particular kind of integration is stored against the integration type By the time client number 50 comes in, the system has already processed 49 versions of “how this kind of build usually breaks in week 6” The loops aren’t autonomous I review every output They don’t make architecture calls. They surface candidates. They draft documentation. They propose feature additions based on what worked on similar prior builds. The human stays in the loop on every decision that touches client value But what they unlock is a kind of leverage that hiring never produced A senior engineer brings 10 years of pattern recognition The loops carry every pattern from every build I’ve ever shipped, queryable in seconds The engineer’s pattern recognition gets better over their career The loops’ get better every Monday Boris is naming this transition at the mode level the shift is exactly the same at the agency level The work is no longer prompting the AI. The work is writing the loops that prompt the AI, and accumulating the data the loops run on Most agencies in 2027 will be one of two things: Either headcount scaling shops still hiring juniors or loop scaling studios with a fraction of the team and 10x the throughput per builder Not fully autonomous yet Not pretending to be either But the architectural shift is real and most operators haven’t internalized that it applies to their business too, not just to the AI building it