The Manual → Chapter 07
Operating in Public
I run my own firm as an operating system: nine AI subsystems doing the back office every day. This chapter is the working log of building it. The mistakes are left in, because that is where the lessons are.
Read in this orderStart with the first 3 below. They set up the rest.
22 essays
07-Aug-2026
01Why I Run My Business as an Operating System Not a CompanyI do not have a team. I have a system. Three businesses run on it. I used to think growth meant headcount. Then I saw every hire was a wrapper around a decision, a rule, a handoff. So I stopped hiring the wrapper and wrote the rule.
13-Jul-202602A Tool Is Not an Operating SystemTools execute steps; an operating system owns loops. Why counting automations tells you nothing, and what an installed operating spine changes.
20-Aug-202603What Founders Ask Before They Trust an AI Operating System for a BusinessFounders stopped asking what my AI can do. They ask what it refuses to do without them, and whether it says when it is unsure. Capability is table stakes now. Predictability is the product
08-Oct-2026The Stamp Said Twelve Checks Ran Two HadYou don't, until you check the checker. A report a system prints about its own work is a claim, not proof, and it stays a claim until something that gains nothing from a pass has verified it. This is the quiet cost of the operator to builder transition: you stop inspecting the work and start inspect
02-Oct-2026Operator to Builder: My Own Report LiedYes, and it happens most to people who build their own systems. A tracker only knows what it was told, not what changed elsewhere. Somewhere in the operator to builder transition you learn to check the work against reality before you trust the record that describes it.
16-Sep-2026Every Business Is the Same Seven BoxesFor most of my working life I was an operator. I ran operations inside large businesses, the kind that turn over hundreds of crores, and I assumed the experience I was building was a stack of industry knowledge I could carry from one place to the next.
21-Aug-2026Every Agent I Build Now Ships With a List of What It Must Not DoThe first version of every agent I built could act. The version I trust can refuse. Writing the do-not list took longer than writing the capability. An agent that cannot refuse is not autonomous. It is unsupervised.
14-Aug-2026The Friction Log Is the Most Valuable File in My SystemI do not keep a file of what my business does well. I keep a file of every place it rubbed. That is the one I open first. Wins tell you what to protect. Friction tells you what to build. #CompanyOfOne #OperatingModel #OperatorToBuilder
12-Aug-2026From Operator to Builder: What I Learned by Deleting Half My AI AgentsI spent eight months fixing every problem by adding another AI agent. Then I deleted half of them in one weekend and the system got better. More workers hid a design problem
05-Aug-2026My Automation Read Nothing and Called It an AnswerLast week my automation built a full report out of nothing and reported success. One wrong character, an empty fetch, a confident wrong file, zero errors. Automation doesn't remove judgment. It relocates it to the moments you forgot to design.
31-Jul-2026Two of My Systems Shared One Resource and Neither KnewA system of mine went silent for two days. Nothing broke. Two jobs polled the same bot. Only one can read at a time. The new one took the stream; the old kept looking healthy and heard nothing. The risk was never the weak part. It was the shared one.
29-Jul-2026My Best Automation Does Nothing Most WeeksMy best automation does nothing most weeks. It reads one number, finds it healthy, and shuts down. Building something that stays silent on purpose was harder than building the thing that runs. Activity was never the point.
24-Jul-2026From Operator to Builder: Why I Let My AI Agents Make Mistakes on PurposeI built my AI agents to be allowed to make mistakes. On purpose.
A reversible error behind a human approval costs me almost nothing. A perfect system I never ship costs me every day.
Speed compounds. Polish doesn't.
22-Jul-2026My AI Rulebook Has a Hard Page LimitMy AI OS runs nine subsystems. Its rulebook is capped at ten pages. Every new rule has to beat an old one to get in. More rules never meant more control. Ranked rules do.
17-Jul-2026The Fix that Created a New Failure!I wrote a rule so my agents would stop skipping a step. They stopped skipping it. They started waiting instead. The patch worked and the system got slower. The skipping was loud. The waiting was quiet. Quiet is worse.
13-Jul-2026The Auditable Org: The Question Nobody Asks Their AI AgentsThe trust problem in agentic AI is not capability, it is provenance. Why decision trails and rollback proof decide whether you can run a business on agents.
13-Jul-2026Stop Counting Deliverables. Count the Hours You Got Back.Deliverables are the wrong measure of operating work. The honest metric for a founder is operating hours freed per week. Here is how to count them.
10-Jul-2026I Do Not Fix a Problem the First Time It HappensI don't fix a problem the first time it happens. The first time is noise. You can't read a signal from a sample of one. First time I log it. Second time I investigate. Third time I build the fix.
08-Jul-2026Operator to Builder: Why My AI Agents Have to Earn Autonomy?No agent in my AI org goes live on day one. Each one earns autonomy: clean dry runs, full observability, a watchdog already watching. The model is rarely the weak link. The missing gate is. Autonomy is earned, not configured.
12-Jun-2026Backup protects your files. Continuity protects your operationTwo partners. No full-time team. The work still ships, because the system runs while we sleep. AI is not our tool. It is our infrastructure. Tools get rented. Infrastructure gets built.
07-Jun-2026A Redirect Can Work Perfectly and Still Be Wrong. The Whole Difference Is One Digit.A 307 and a 308 land a visitor on exactly the same page. To every machine reading your site, they mean opposite things. On the layer you never test, and why it is getting more expensive to ignore.
03-Jun-2026What My MBA Missed About the Operator to Builder TransitionThe MBA taught me to name problems. The Rs.400 Cr factory taught me to feel them before they had a name. One makes you articulate. The other makes you right.
If you want this running for you
Start with one working system built around your week.
The AI Setup Sprint is two or three focused sessions that leave you with a Claude-based system doing real work. Fixed price. No ongoing commitment.