I transform companies into AI organizations for a living.
That work has two halves. The first is what your board cares about: operating model, governance, adoption, measurable change across engineering and the functions that were never supposed to need a terminal. The second is what I had to learn the hard way: how to build the software agentically so the transformation is real, not a slide deck. How to help individuals redesign their work so it is actually AI-native, not “we bought ChatGPT seats.”
I have spent the last year on both. The tooling became AIOS. The uncomfortable personal discovery is this: the better I got at agents, the worse I got at focus.
If you are a CEO or founder who studies AI constantly, ships experiments every week, and still feels like you lost something important along the way, you probably know the feeling. You used to have flow. Long stretches where the work was hard and you were fully in it. Now you have tabs. Notifications. A dozen half-finished threads and the nagging sense that you are the bottleneck in your own company.
This post is not another “top ten prompts” piece. It is the operating rhythm I am using to get flow back while running agents at scale. AIOS is my experiment in reclaiming deep work without pretending the fleet does not exist.
You can be good at AI and bad at attention
There is a maturity model underneath this (agentic maturity). It scores how well you delegate, verify, compound learnings, and govern cost. You can climb to Spine L4: Agentic Engineering and still fail the thing your people actually need from you: judgment, presence, strategy, the conversation only you can have.
The signal I watch now is not another axis on a rubric. It is the Attention card: context switches per hour, how long your focus blocks last, how many agent sessions you are juggling at once. When that reading says orchestration-heavy, it means you are not building anymore. You are scheduling.
That is not a character flaw. It is what happens when every agent interruption is treated like urgent. I built the asks queue and the cognitive ergonomics research lens because I needed language for it. Transformation work needs a human who can think. Not a very expensive notification handler.
One command to start the day
npm run aios -- loop dailyThat is the morning brief. What changed overnight. What is blocked. What you owe today. At the top, if anything is waiting: Attention (blockers) and Queued asks (decisions and FYIs).
The asks queue is the design bet in one data structure. Agents do not stop the world with a modal when they need you. They enqueue. You drain on your schedule.
No lost escalations. No fake urgency. No approving the third draft of an email you half-read because you were mid-thought on something that actually matters.
The whole day
A founder's day in five phases: morning orientation with loop daily and pipeline review; a deep work block for human-only revenue work; an orchestration block with up to five agents drafting into the asks queue; a second deep work block for calls and decisions; and an evening compound step for notes, artifacts, and draining the asks queue.
Read it top to bottom. This is the rhythm I am trying to live.
Morning is orient: loop daily, a short pipeline check (who needs you today,
no more than five), pick the mode.
Deep work block one is ninety minutes of human-only work. The strategy memo. The case study. Prep for the board conversation or the client discovery call where your judgment is the product. One session. Notifications off.
Orchestration is the fleet: research drafts, content variants, internal briefs queued into asks. Capped at five concurrent top-level agents. You review and approve. You do not live in the tabs.
Deep work block two is calls, negotiation, being in the room.
Evening is compound: what you learned today becomes an artifact (a rule, a
template, a skill step) so tomorrow’s agents start smarter. Then aios asks drain for inbox zero.
Two modes, one switch
Most founders do not need more agents. They need a clean boundary between thinking and supervising.
aios mode deep-work # local ping offaios mode orchestration # ping back onTwo attention modes toggled by aios mode: deep work silences the local notification ping for focused human-only work in one session; orchestration restores the ping for supervising up to five concurrent agents whose drafts land in the asks queue.
Deep work means one session, one thread, one outcome.
Orchestration means the fleet can ping you when a sub-agent finishes. Fine for drafting options you will never send all of. Fine for background research you will skim at a boundary. Not fine when you are writing the paragraph that sets company direction.
The toggle flips one setting in Claude Code. It remembers what you had before. Deliberately small so you actually use it.
How asks fit in
Old world: agent hits a decision, everything stops, you context-switch, you half-read, you approve, everyone pretends that was governance.
AIOS world:
| Severity | Meaning | When you see it |
|---|---|---|
blocker | Something is stuck | Top of loop daily · drain first |
decision | Needs your call | Queued asks · batch at boundaries |
fyi | Done, review when ready | Same queue · no fake urgency |
Twice a day, run a drain boundary: list, resolve, move on. End of day:
aios asks drain.
Agents keep running on safe defaults where they can. You stay in deep work until you choose to orchestrate.
Measure the human, not just the fleet
Once a week:
npm run aios -- analyze --since 7dFive axes for engineering maturity. An Attention card for your rhythm. A cognitive ergonomics shadow score compared to your own baseline, not anyone else’s.
The plan is not “use fewer agents.” It is:
- Compound learnings (every correction becomes a rule, skill, or template)
- Cap concurrency (five top-level sessions, not fifteen)
- Protect two deep blocks (the work only you can do)
- Let the fleet draft (orchestration is for preparation, not substitution)
Transformation leadership is still leadership. The model should make the human more present, not less.
A loop is still a loop
Everything above sits on the same foundation as loop engineering: a task plus a check. The operator layer does not replace that. It makes sure you are not the check on every low-stakes draft at eleven at night.
A loop is a task plus a check. The cycle runs clockwise: you delegate a task, an agent acts on it, a check decides whether the work is good enough to stop, wrong enough to retry, or stuck enough to escalate, and the loop adjusts and runs again. The check is the emphasized part, because a task without a check is just hope.
Delegation without a check is hope. Orchestration without boundaries is burnout. The mature move is a longer leash on low-risk work behind a real verifier, and a short leash on your own attention.
What I am optimizing for
Maturity is the human regaining deep work, not answering agents all day.
That sentence is in the cognitive ergonomics spec because I wrote it before I believed it. I believe it now. The companies that win this transition will not be the ones with the most agents running. They will be the ones where leaders and teams can still do the hard thinking, and agents handle the verifiable middle.
If you are building with AI every day and mourning the focus you used to have, this is the system I am dogfooding:
loop dailyto orientmode deep-workwhen the work is yours alonemode orchestrationwhen the fleet draftsasksso nothing blocksanalyzeso you cannot lie to yourself- compound so next week starts smarter than this one
The transformation is the company. The rhythm is you. Both have to hold.
Further reading