# The three-person company that operates like fifteen > AI agents now do real work across real businesses. Here is how we build that way at the studio, and a practical setup you can start this month. URL: https://regenerate.capital/insights/three-person-company/ Language: en · Other language: https://regenerate.capital/fr/analyses/three-person-company/index.md Publisher: Regenerate Capital (https://regenerate.capital) · Updated: 2026-09-25 --- [Home](https://regenerate.capital/) / [Insights](https://regenerate.capital/insights/) Thesis · 6 min read # The three-person company that operates like fifteen AI agents now do real work across real businesses. Here is how we build that way at the studio, and a practical setup you can start this month. Joaquim Miro · 2026-09-25 A team of three can now ship and operate what once needed fifteen people. I don’t say that as a prediction. I say it because I run it that way every day, across my own companies. Agents do research, outreach, operations and reporting for us. Some have run on a schedule for months, and I will be honest about where they break. But the shape of a company has changed, and founders who build this way now will have a real head start. ## What does “real work” look like? Two examples from my own builds. The first is Stroby’s outreach agent, called Echo. It finds high-signal newsletter creators, drafts personal replies in our voice and sends them. I built it myself, with no team, and it went live in under a month, for under $1,000 in total build cost. The second is a personal job pipeline I built to test the same idea on my own career. It scrapes about 450 candidate roles per run from 55 sources, scores them with one model, analyzes the best ones with a stronger model, and tracks responses by email. It runs on a schedule, inside a strict monthly budget, with a kill switch and daily caps. Neither of these is a demo. Each one replaced hours of work a person used to do by hand. ## The model is the cheap part Here is the honest dose of reality. The hardest part of an agent is rarely the language model. It is everything wrapped around it: logins, sending, scheduling, safety filters, budgets and the moment a human needs to step in. On Echo, I spent more time on the safety layer than on the prompts. It has intensity gating, filters against generic filler, suppression rules and deterministic checks as a backstop, because prompt rules alone don’t hold at volume. Code does. That is why every build at the studio starts with one question: which business number does this move? If we can’t answer it, we don’t build the agent. ## How we build this way at the studio Every studio company inherits the same operating stack from day one. Gatewayz.ai, which I co-founded, handles inference and routes across hundreds of models, so each job can use the right model at the right price. On top of that, each company runs its own small fleet of agents. When agents act for a company, someone has to answer for them. That is the idea behind the agentic autonomous organization, or AAO: humans and agents share the work under one rulebook. Studio company FlashyOS publishes an open specification for it, and SafeCommit provides signed checks before an organization proceeds with a counterparty. ## A practical setup you can start this month You don’t need a fleet to begin. You need one job, done well. Here is the sequence I use: - Pick one recurring job that eats at least five hours a week. Lead research, weekly reporting and inbox triage are good first candidates. - Write down the number it should move and how you will measure it. Hours saved, replies booked or errors caught. - Run it in draft-only mode for a week. The agent prepares the work and a human approves every output. On my job pipeline, I ran seven days of drafts before anything went live. - Add hard limits in code: a daily cap, a budget ceiling and a kill switch you can flip in one command. - Name one human who answers for the agent. If nobody owns it, it will quietly break. - Review it weekly. Read a sample of outputs, check the number, and tighten the rules. Once the first agent is trusted, add the second. Most of the value comes from the third and fourth, when the agents start feeding each other. ## What still needs people? Judgement, relationships and taste. An agent can draft a hundred outreach messages. It can’t decide which partnership will matter in three years, or sit across from a founder having a hard week. The future of a senior job, in my view, is being a great delegator. The goal is not fewer people doing more busywork. It is a small team spending its time on the decisions only people can make. ## Where to start If you are running a small team and want to find your first agent, write down the three most repetitive jobs in your week. Then send them to me through the studio. I will tell you which one I would automate first, and how I would keep it safe. Onwards and upwards. Subscribe ## Get the next note. Work with the studio ## Building something worth growing? Tell us what you’re building and what you need. We read every application, and every founder gets an answer. [Apply to work together](https://regenerate.capital/apply/)