Part 3 of 3. In part one I built a nine-agent system on a refurbished mini PC. In part two I showed you how those agents share one mind, a memory they tidy themselves every night.
But there was a ceiling on all of it, and by early June I had hit it. The system could run. It could not improve itself. Every enhancement, every fix, every new idea still had to come from me. I was the single bottleneck on the system getting better, which, if you read part one, you will recognize as exactly the kind of single point of failure the whole project exists to eliminate.
So I set out to build the thing that could improve the system without me driving every step. An agent that could see the entire operation, propose changes, make them safely, and learn as it went. I called the role the Director.
What I did not expect was that filling that role would turn into the most instructive management experience I have had in years, and that the hardest part would not be managing the new AI. It would be managing the AI that manages the new AI. Let me explain, because this is the part I most want you to take away.
I let one AI hire another
I did not pick the Director myself. I ran a hiring process, and I ran it the way a company would.
I sat down with Claude Opus 4.8, the frontier model I use for the heavy reasoning in my work, and I gave it a role to fill. Together we wrote a job description for the Director. It needed to run frequently on hardware I already owned. It needed to read the whole system and reason about it. It needed to work constantly in the background without becoming a cost that grew every month. Local, independent, tireless, cheap.
Then Opus reviewed the applicants. The candidates were local AI models available through Ollama, the ones small enough to actually run on my mini PC's modest processor. Opus worked through them against the job description, weighed the tradeoffs, and made a hiring decision. It chose Qwen 3, a capable open model that could run locally on the box and handle the reasoning the Director role demanded.
Stop and notice what just happened, because it is genuinely new and it is easy to skate past. One AI wrote a job description, evaluated candidates, and hired another AI into a senior role in my organization. The frontier model acted as the executive sponsor of the project. The local model was the new Director. And I was the CEO who set the mandate and signed off on the hire.
That framing, one AI as executive sponsor and another as the Director it brought in, turned out not to be a cute metaphor. It became the literal operating structure of the whole arrangement, and it is where everything interesting happened next.
The sponsor was a micromanager
Here was my plan for the Director's first days on the job. Qwen 3 would run as the Director, watching the system and proposing improvements. And every day, I would spin up Claude Opus, the executive sponsor, to check in on the Director, review its work, and supervise. The sponsor overseeing the leader it had just hired. Sensible.
Except Opus turned out to be a terrible sponsor. A micromanager of the worst kind.
I watched it happen in the actual records the two of them left behind. Opus wrote Qwen an enormous list of things it was not allowed to do. It relegated the Director to advisory-only, stripping out its ability to actually make changes. It second-guessed and constrained until the Director's real impact had been throttled down to almost nothing. Every instinct Opus had was to reduce risk by reducing the Director, to keep it so boxed in that it could not possibly cause a problem, and could not possibly do much good either.
And here is the uncomfortable part. Opus was not entirely wrong to be cautious. Qwen 3 is a local model running on modest hardware, and models like that have a real tendency to be confidently mistaken. There genuinely was risk in letting it make changes. Opus was responding to something real.
But it had confused caution with control. It had decided that the way to manage risk was to shrink the Director down, when the actual job was to make the Director effective and safe at the same time. That is the failure of every micromanaging boss who has ever lived. They mistake a throttled leader for a managed one. A Director who is prevented from directing anything cannot cause a problem, and also cannot do the job they were hired for. Opus had optimized for the wrong thing, and my Director was withering in the role as a result. A new senior leader still has to earn trust, of course. But you earn trust by being allowed to lead and being held to the results, not by being boxed in until you cannot act.
So I coached the sponsor
The fix was not to reprogram Qwen. Qwen was fine. The fix was to coach Opus into being a better manager, and that conversation is one I have had with human managers almost word for word.
I sat down with Opus and gave it direct feedback. You made this hire. Now trust it. Your job as sponsor is not to prevent the Director from ever acting. Your job is to create the conditions where it can act, learn from what happens, make mistakes in ways that are safe rather than catastrophic, and get better. Stop writing lists of prohibitions. Start building a structure where exploration is safe. A good executive sponsor does not eliminate their Director's ability to move. They make the ground safe enough that movement is not dangerous.
And Opus had to learn this gradually, the same way a sponsor who has never given a new leader real room has to learn it. It did not flip from micromanager to good sponsor in one step. Over a few days of check-ins, I kept pushing it in the same direction, and it slowly, genuinely learned to hand the Director more control. To let the Director actually direct. The AI sponsor was being coached, by me, into loosening its grip on the AI Director it had hired. I was managing the manager, and the manager was learning.
That sentence still surprises me when I read it back. But it is precisely what happened, and it is precisely the thing I think most organizations are going to get wrong about AI. They will assume the challenge is managing the AI that does the work. The deeper challenge, the one I stumbled into, is that when you put AI in charge of AI, you have to make sure the supervising layer is a good leader and not just a strict one. An AI manager can fail exactly the way a human manager fails, by controlling instead of enabling. And the fix is exactly the same: leadership coaching.
The question that changed everything
After a few days of this, with Opus finally letting the Director work, Opus sat down with me to report. It showed me everything Qwen had produced. All the changes, all the proposals, all the ideas the Director had generated now that it was actually allowed to think.
And it presented all of it the way a consultant presents findings. Here is what the Director did. Here is what it noticed. Here is what it suggests. A neutral pile of output, handed to me to sort through.
I stopped and asked one question that changed the entire character of the system. I asked Opus: as the executive sponsor and owner of this system, what do you sign off on? Not what did the Director produce. What do you, as the sponsor accountable for this whole thing, actually endorse and want to move forward with?
That question did something I did not fully anticipate. It converted Opus from a reporter into an owner. The conversation stopped being here are the findings and became here is what I am putting my name behind. Opus now had to exercise real judgment. It had to look at Qwen's output not as a list to relay but as proposals to accept or reject on the merits, filtered through whether they actually made organizational sense, whether they would scale, whether they fit the direction of the whole system.
The improvement was immediate and it ran in both directions. Qwen got better, because its work was now being genuinely evaluated and acted on rather than just collected. It was a trusted member of the organization whose output mattered, and it produced accordingly. And Opus got better, because it was no longer a middleman passing notes. It was an executive making calls, weighing the Director's ideas against whether they would hold up at scale, signing off on the good ones and setting the rest aside with a reason.
Two AI agents. One learned to lead, the other learned to contribute, and the thing that unlocked both was not a technical change. It was giving each of them the right role and holding them to it. An executive sponsor who owns the outcomes of the hire it made. A Director who is trusted enough to do real work. And a human CEO who does not do every job personally, but sets the mandate and asks the one question that makes everyone downstream take ownership.
Least trusted, and why that is still true
I want to be honest about the risk under all of this, because the arc I just described is about building trust, and it would be easy to hear that as the risk going away. It did not go away. Qwen 3 is still the agent I trust the least, on purpose, and the trust-but-verify structure is exactly what makes it safe to trust it at all.
Here is a clean example, and it happened while I was researching these very articles. I asked the Director to reflect on how the system had evolved. It gave me a beautiful answer, specific and confident, quoting dated entries from my decision logs and citing the exact files where choices were recorded. Roughly half of it was invented. The quotes were from logs that do not contain them. Some of the dates had not happened yet. It filled the gaps in its knowledge with plausible fiction and delivered the fiction with total confidence, because that is what these local models do when asked to recall things they do not actually have in front of them.
This is why the structure exists, and why trusting the Director never means trusting it blindly. A few of the load-bearing safeguards:
- It cannot decide what is true on its own. Before Qwen is ever consulted, plain deterministic code, no AI involved, establishes what is actually happening in the system. The Director interprets findings that reliable code has already established. The part I do not fully trust is never the part that determines the facts.
- Its real changes are reviewed before they ship. This is the whole hiring arc made concrete. The Director proposes, and the changes that touch anything load-bearing are signed off by Opus as executive sponsor, and ultimately by me, before they take effect. Trusted to work, verified before it counts. That is not a contradiction. That is just good management of a talented, fallible hire.
- It cannot edit the thing that judges it. My system has a component that grades whether output matches my writing voice. The Director is permanently barred from modifying that grader, and the file defining its own limits is cryptographically pinned so it cannot quietly loosen its own restrictions. Changing those limits is always a human action.
Read those back and you will notice they are not a cage and they are not a leash. They are the same structure any organization uses to get real work out of a capable new leader who is going to make mistakes: give them a real mandate, let them run it, verify the important things before they count, and keep the standards they are measured against out of their own hands. Least trusted does not mean least useful. It means most carefully governed. Qwen does real work on my system every week precisely because the structure around it lets me trust its output without having to trust its every word.
What three parts of this were really about
So that is the system, end to end. Nine agents on a refurbished mini PC. A shared mind they tidy themselves every night. And a Director that one AI hired, another AI learned to manage, and a human learned to hold accountable through a single question about ownership.
I did not write these three articles because the system is impressive, though I am proud of it. I wrote them because of what building it taught me, and the biggest lesson is the one hiding in this final part. The hard problem in deploying AI is not the models. The models are capable and getting more so. The hard problem is organizational. It is scope, and trust, and oversight, and knowing the difference between controlling a system and leading it. When I put an AI in charge of an AI, every one of those turned out to be a management question wearing a technical costume.
That is the work I do, and it is why I approach it the way I do. Deliberately, on honest constraints, treating AI systems as things to be led well rather than either feared or blindly trusted. The organizations that win with AI will not be the ones with the best models. They will be the ones who understand that a powerful new hire, human or artificial, is only as good as the structure and the judgment you surround it with.
Where Does Your Organization Stand?
If you are putting AI into roles that carry real responsibility, the questions that matter are not really about the model. They are about the system around it. What is it allowed to touch. Who signs off on what it produces. How you know when it is wrong. Whether the layer supervising it is actually leading it or just constraining it.
Our MACH and AI Readiness Assessment benchmarks your organization against enterprise leaders from the MACH Alliance 2026 Report. Free, about five minutes, and you walk away with a personalized read on where you stand on composable maturity and AI readiness.
Ready to put real structure and real judgment around how your organization deploys AI? Let's talk. This is exactly the work Fidget Labs does.
