--- title: "Bro Talks: Rafaela Ostrowski on Running a Nine-Person Studio Like a Forty-Person One" dek: "The founder of Northpaw Studio ships client software with nine people and a bench of agents. She talks pricing, hiring, taste and the day an agent nearly lost them their biggest client." author: "Theo Brooks" author_title: "Staff Writer, Startups" category: "Startups & Ideas" series: "Bro Talks" format: interview tags: ["Work", "Agents", "Coding", "Startups"] published: 2026-11-04T06:37:07.284Z updated: 2026-11-04T06:37:07.284Z reading_minutes: 5 url: "http://localhost:3000/article/bro-talks-rafaela-ostrowski-northpaw-studio" ai_disclosure: human ai_disclosure_label: "Human-written" members_only: false summary: - "Northpaw organises its work in three-person pods (lead, engineer and designer), each backed by agents set up for that client." - "A wall of fourteen 'things agents may not do' rules, each one learned from a mistake, governs the studio." - "The studio prices by outcome rather than by the hour, so it keeps the gains from working faster." - "Junior engineers regularly build features by hand, both to learn the craft and to learn where machine-written code goes wrong." publisher: "Ideabro" --- # Bro Talks: Rafaela Ostrowski on Running a Nine-Person Studio Like a Forty-Person One *The founder of Northpaw Studio ships client software with nine people and a bench of agents. She talks pricing, hiring, taste and the day an agent nearly lost them their biggest client.* Northpaw Studio builds software for other companies: booking systems for physiotherapy chains, internal dashboards for a regional grocery wholesaler, a quoting tool for a family-run scaffolding firm. It has nine employees. Its founder, Rafaela Ostrowski, says it delivers the volume of work her last agency needed forty people for. The trick, she insists, isn't about squeezing headcount. It's a different way of organising a studio. Small human teams write the specs, review relentlessly and own the client relationship. Agents back them up with first drafts of code, tests, documentation and project admin. We talked in the studio's converted print shop, where a whiteboard titled "Things agents may not do" takes up an entire wall. *This interview has been edited for length and clarity.* ## On building a different kind of studio **What made you start Northpaw instead of just using AI at a regular agency?** I tried at a regular agency. The problem wasn't the tools. It was the org chart. Everything was designed around handoffs between people: the designer hands to the developer, who hands to QA, who hands it back. When an agent can do a first pass of three of those steps in an hour, the handoffs become the slow part. You can't bolt that onto an existing structure. You have to start from the handoffs. **So what does a Northpaw team look like?** Pods of three. There's a lead who owns the client and the spec, an engineer who owns architecture and review, and a designer who owns taste, and I mean that word seriously. Each pod has a set of agents configured for that client. One drafts code against the spec, one writes and runs tests, one keeps the documentation and changelog current, and one writes the Friday update email, which a human always edits. **How much of the code do agents write?** First drafts? Most of it. Final code? That's the wrong question. Every line gets reviewed by an engineer who could have written it. We aren't hiring people who can't code and pointing them at a machine. We hire very good engineers and ask them to spend their time on the 20% that's hard. ## On the wall **Tell me about the "Things agents may not do" wall.** It started as a joke and became our most important document. There are fourteen rules now. No agent emails a client. No agent touches production data. No agent merges its own work. No agent decides a deadline. **Why deadlines?** Because a deadline is a promise, and only a person can make a promise. An agent will estimate optimistically every single time, and it will never be the one on the phone apologising. > Every rule on that wall is a scar. We don't write a rule until something has gone wrong, or nearly gone wrong, at least once. **Which rule has the best story?** Number nine: no agent changes a dependency without a human reading the changelog. Last year an agent upgraded a library to fix a small bug. The upgrade was correct, but the new version changed how dates were formatted, and that was a disaster. Our biggest client's appointment reminders started going out with the day and month swapped. For about four hours, patients in one city were told to show up in the wrong month. **How did you handle it?** I called them before they called us. We'd caught it from a customer complaint and rolled it back, and I had a timeline ready within the hour because our agents log everything they do. The client said the logs were the reason they stayed. That's when I understood the audit trail is part of what we sell. ## On pricing and money **How do you charge, if the work takes fewer hours?** Not by the hour. That model punishes us for getting faster. We price by outcome: a fixed fee for a defined piece of software, plus a monthly fee for maintenance and improvements. Clients like it because they know the number. We like it because the efficiency gains are ours to keep. **Are clients nervous when they hear agents are involved?** Some are. We tell everyone up front, in the contract, with a page explaining exactly what agents do and what they never do. The nervous ones usually relax when they see the wall. The ones who don't relax aren't our clients, and that's fine. **What does all this cost to run?** Model and tool costs are our third-biggest expense, after salaries and rent. It's real money, roughly what we'd pay one junior employee. But it's predictable, and it grows with the amount of work rather than with time. ## On hiring and the next generation **Here's the worry everyone has: if agents write the first drafts, how do juniors learn?** That's the question I lose sleep over. Our answer, for now, is an apprenticeship rule. Every month, our two most junior engineers write certain features by hand, without agents. Then they compare their version with an agent's draft and present the differences to the team. They learn to code, and they learn what machine-written code gets wrong, which might be the more valuable skill. **What do you look for when you hire?** Writing. I know that sounds strange for a software studio, but a spec is writing, a review comment is writing and a client update is writing. If you can explain clearly what "done" means, you can direct agents and people equally well. If you can't, no tool will save you. **And taste?** Taste is noticing that something is technically correct and still wrong. The button works, but it's in the place where nobody looks. The email is accurate, but it sounds like a parking ticket. Agents are very good at correct. They're mediocre at "still wrong," and that gap is where our designers earn their salary. ## On what comes next **Will Northpaw stay at nine people?** I want us to stay small enough that everyone still knows every client's name. Maybe fifteen people. We'll get more capable before we get bigger. **What's the most common mistake you see other studios make?** They treat agents like cheap juniors instead of tools that need a workshop. They give them vague tasks, no rules and no logs, and then they're shocked when the results are vague, unruly and impossible to audit. The agents are the easy part. Building the workshop is the job. **Last question. What's the next rule going on the wall?** We're debating it now. Something like: no agent writes the first sentence of anything a client will read. The first sentence sets the tone, and people should own the tone. --- Source: [Bro Talks: Rafaela Ostrowski on Running a Nine-Person Studio Like a Forty-Person One](http://localhost:3000/article/bro-talks-rafaela-ostrowski-northpaw-studio) — Ideabro. Human-written: Reported and written by a person. AI not used for the text.