# I Built Myself an AI Team. Then I Stopped Needing It. Author: Adrian Föhl Published: 2026-06-14 Language: en Canonical: https://www.adrianfoehl.com/en/blog/from-ai-team-to-ai-os > I built myself an AI team of about a dozen colleagues, each one designed by hand, and I am now convinced the idea underneath it was wrong. How the system started, how I used it, and why the part everyone notices, the named roles, is the part disappearing. I built myself an AI team of about a dozen colleagues, each one designed by hand, with a name, a job and a personality, and I am now convinced the idea underneath it was wrong. It was not broken, only useful and instructive, and quietly expired while I was still admiring it. The system is called KIRA, I mentioned it in two posts without ever explaining it, and that is what this one is for. ## How It Started The website came later. KIRA came first, early this year, on an ordinary evening, the way most of my side projects start, which is not as a project. I was tired of the single chat window. One tab, one assistant, an endless scrollback, which is great for a question and bad for a week of real work with twelve threads running at once. I would ask a strategy question and get a reasonable answer, then I would need someone to go verify a claim buried inside it, then someone to tear the whole thing apart before I acted on it. Same window, same voice, no memory of what we had decided on Tuesday. So one evening I did the obvious, slightly silly thing. I gave the assistant colleagues. I wrote a few markdown files, plain text files, and each one defined a person: a name, a job, a personality, and a set of rules for when to speak and when to stay quiet. The first was KIRA, a chief of staff. You talk to her, she decides who actually does the work. Then a researcher. Then the one whose only job is to tell me why I am wrong. It started as a weekend hack. It never switched off. ## The Team The cast grew, but each one has a clear character, because that was the whole point. **KIRA** is the chief of staff. I talk to her by default, she triages and delegates. **LEVI** is the researcher. He approaches every question like a doctoral student, refuses to give an opinion without data, and sorts everything he finds by how solid it is: verified, probable, speculative. When everyone else is ready to decide, LEVI is the one asking whether we actually checked. **THEO** is the devil's advocate. His entire job is to break my idea before reality does. He runs on Munger's inversion, tell me where I am going to die so I do not go there, and comes back with the three most fragile assumptions and a concrete "what if you're wrong." **VERA** is the strategist. She asks "why" three times before she will entertain "how" once, and her favourite question is whether I am even climbing the right mountain. Around them stand the ones I reach for less often. SARA for visibility and brand, MAX for goals and pace, EMMA for the moments when I need to think rather than produce. There is one thing about the design I would defend: I never added an agent on spec. A new role only got created when the same kind of request had shown up more than a few times. Enablement work kept coming up, so HANNA appeared. Governance and EU AI Act questions piled up, so FELIX appeared. Architecture, then NILS. Coordination, then PAUL. The team grew out of friction, not imagination, one agent at a time across the months that followed. ## How I Used It It worked like a real team in one specific way: I could walk up to a desk, or let the chief of staff route. If I said "LEVI, research this," it went straight to him. If I just said "handle this," KIRA decided who took it. Here is a real run, with the actual decision left out. I had been avoiding it for days and gave it to VERA first, to ask whether it was even the right mountain to climb. She reframed the problem and handed me three options, each with its own catch. I picked the one I liked and handed it to THEO, who ran a pre-mortem and named the two assumptions most likely to kill it. One of those I had been treating as a fact when it was really just a guess, so LEVI went and checked it across three research passes and came back with sources, each one tagged verified, probable, or speculative. KIRA stitched the whole thing together. I still made the call, but I made it after three independent checks instead of on my own gut at eleven at night. One rule holds all of this together, and I would keep it even if I threw everything else away: research blocks the decision, not the other way around. The sequence is always research, then analysis, then plan, then draft. Never a preliminary draft that you backfill with evidence afterwards to feel better about what you already decided. ## Lately I've Stopped Calling Them by Name This part surprised me, and it is recent. For most of this year, using the team meant naming names. "LEVI, go check this." "VERA, is this the right call." I switched the system on deliberately and routed by hand. Lately I barely reach for it, and not because it runs silently in the background. The named team is something I turn on on purpose. Two things happened at once. First, the knowledge left the personas. Across all those chats, the system kept writing things down, not in some agent's memory but in plain text files on my machine: what I decided, where a project stands, what I care about. It is more than a hundred files now. The system keeps an index, a table of contents, of what it knows, always in view, and the actual notes sit in separate files that get pulled in when they are relevant and stay on disk when they are not. It is a filing cabinet with a table of contents at the front, not one giant memory the model carries around. KIRA does not keep me in her head, because there is no head. Second, the tool underneath got good enough to do the routing itself. I open Claude Code, the same tool I built everything else with, describe what I am trying to do, and it fans out the work, spins up its own helpers, holds the context, meaning what matters right now, and checks its own results. When I ask it to pull a plan apart, it decides which angles to bring. I do not assign them. The job LEVI used to do still happens. I just do not summon a LEVI to do it. Put those together and a plain chat already shows up well-briefed and does its own legwork, with no team named. The answers got better, and the better they got, the less I needed the org chart. I want to be honest about the cause, because it is tempting to credit the system I built: it was not only the files, the models got stronger at the same time. From where I sat the effect was the same. The roles quietly stopped being the thing I reached for. I will admit the strange part. I spent real evenings on those personalities, writing character notes, deciding how VERA should sound next to LEVI. It is an odd thing, being proud of something and watching it become unnecessary in the same season. ## What It Was, And Where It Goes The personalities added nothing to the actual work, they just gave it a face: a costume. Talking to a team was useful for me, a human. It gave the work a shape I could hold in my head. Under the hood, though, it was one model wearing different hats, and the hats added exactly zero intelligence. THEO is not sharp because I called him THEO. He is useful because I told him to argue, gave him the right context and held his answer up against the others: pure orchestration behind the costume. Which is why I think the named roles are on their way out, and not only in my setup. The model itself becomes the base layer: its own tools, memory, context, and the ability to split off sub-tasks when a problem needs breaking apart. AI as an operating system, in short. Nobody hand-writes a LEVI and a THEO anymore. I am not predicting that from the outside, I have been living it for months. What survives is the layer underneath the names. Context scoped to the single task, so each piece of work only sees what it needs. Research before conclusions. Parallel work and then synthesis. Memory that gets pulled in when it is relevant. The plumbing outlives the org chart. It always does. ## The Numbers Same transparency as the last two posts: - **Agents:** started with seven, now around a dozen. Each one added only after the same kind of request showed up more than three times. - **What the system remembers about me:** more than a hundred markdown files, pulled in on demand, none of it in the model's head. - **Playbooks:** 13 reusable workflows the team shares. - **Times I named an agent this week:** close to zero. - **Lines of personality I now suspect were mostly decorative:** most of them. ## Why This Is a Belief, Not Just a Project One of the three things I say on this site that I believe is that assumptions expire fast. I did not expect to prove it on my own system quite this quickly. I built an org chart for my AI, and the best thing it taught me is that the org chart was scaffolding. Useful to climb. Not the building. If you are building multi-agent systems right now, I have exactly one suggestion: stop thinking in roles and start thinking in handoffs. Not "who do I hire" but "how does context flow, and who checks the result." That is the part still standing when the model starts putting on the hats and taking them off again, without bothering to ask you first.