I have killed my single biggest bottleneck three times this year. Every time I killed it, it just relocated. The third one blindsided me. The bottleneck is now my own attention. Here is the shape of it. Phase one: I was the bottleneck. So I moved myself off the critical path. I stopped doing the mechanical work and started directing it. I plan, I decide, I review. Execution runs in the background while I keep thinking in one window. The only thing allowed to interrupt me is a worker that hits real ambiguity and needs a human call. Phase two: The expensive model became the bottleneck. Lean on one premium brain for everything and it turns into a funnel. Everything queues behind it. So I swapped the one costly brain for a pool of cheap and free workers, spread across a mesh of machines. I own the labour and rent only the judgement. One fast worker on one box is still just a queue. Real parallelism lives across the mesh, not inside any single model. Phase three, the one I am in now: A dozen workers run in the background across several machines, and I cannot see them. You cannot direct what you cannot see. So the build became the instrument itself. I stopped building a dashboard and started building a control plane. A monitor shows you the fire. A control plane lets you put it out without leaving the chair. One surface that does not just watch the fleet, it steers it. Spawn, resume, redirect, kill, all from the same seat. The honest limit, because there always is one. This is not self-driving and nothing builds itself. I still review everything that matters. The proof is in how much steering it actually does. I built it as a three-rung autonomy ladder. 1. Bare shell. I watch, and I drive everything by hand. 2. Semi-auto. A stalled or crashed worker raises its hand, and I clear it with one key. Resume, escalate, or kill. 3. Self-driving. It resumes on its own, retries twice, reroutes the work, and only comes to me when it is stuck. The point is that it acts on the workers. It does not just display them.

I built a skill package that gives each AI agent a distinct thinking style, and upgrades the council skill into a forum of truly independent thinkers. The result: agents that think differently on purpose - tailored to task - for sharper, well-rounded output. First, I want to credit @Curtis Hays and @Brooke Hays for the inspiration. If you haven't watched their video in the post "If Your Specialized Agents Don't Think Differently. They Should." - please do. This is their work, I just wrote some files. They touched on something I've been trying to crack since I started messing with agents a few months ago: how do I make an agent think in specialized ways? How can I get an agent to think from first-principles? How can an agent see things in ways that no one else would? How can an agent connect the dots that I am missing? I tried assigning personalities, pointing agents at a knowledge base, etc, etc. Sometimes this worked - but often, they would fall back to whatever llm baseline was dictating their behavior. Brooke's breakdown of cognitive functions, and Curtis' brilliant idea to assign them to his agents was the spark I needed. I started by re-acquainting myself with the personality types, and building a COMPENDIUM: A reference guide to the 8 Jungian cognitive functions, the 16 MBTI types and their full stacks, and how each maps to an AI agent role. It was built from Carl Jung and Isabel Briggs-Myers' work. It's the single source of truth the other two skills read from — the "textbook" behind the system. The result: a shared, verified vocabulary for giving any agent or task a defined thinking style. Then I got to work on figuring out how I could incorporate this into my ICM workspaces. This led to making the COGNASSIGN skill: A skill that assigns the right cognitive "Mind" to one agent or task. You answer a few job-first questions (what must it do, perceive, judge, and be best at), and it picks a lead function plus a balanced partner, then writes a small procedural block telling the agent how to think — not a personality label. If a job needs too many strengths it tells you to split it into two agents; if it's pure mechanical work it returns "N/A." The result: any agent gets a fit-for-task wiring in one drop-in block, the way the video's two agents thought differently from just four letters.