Field Notes from the Agentic Operator - dispatches on building and operating real businesses with AI agents, and what holds up once a demo has to survive a Tuesday.
The new unit of software work is not the task - it is the operating system, and the new job is operator.
Match the enforcement mechanism to the failure mode: encode the irreversible in code, write down the reversible in prose, and never confuse the two.
An audit of my own operator setup revealed the dangerous combination, and why the only real defense against prompt injection is architectural, not a filter.
Three of my scheduled jobs ran broken for weeks with no alarm and no trace; here is why silence is the most dangerous signal in any autonomous system.
Autonomous agents only save time if you stop reviewing every action - so the real unlock is a ledger and tiered trust, not a smarter filter.
A password leaked from a tiny side project taught me that real security comes from self-installing guards, not policies people have to remember.
Why a scatter of email addresses across a small team quietly erodes trust, and why consolidating early is a coherence feature, not a cleanup chore.
Three times I turned an advisory hour into unpaid operator work, and what it cost the only thing my portfolio actually runs on.
A founder team’s months-long stalemate taught me that alignment is not shared belief - it is a shared agreement about what evidence will settle the bet.
The blockers that kill adoption live in the workflow and the definitions, not the model, and they only surface when one real operator tries the job.
A wedge is not a narrow customer - it is the delivery channel that teaches you most about the problem from real usage.
A financial model is not a forecast or a plan input - it is a decision instrument you build to force clarity on one specific choice.
A knowledge base broke past sixty documents because it was organized by where things got used instead of what they actually were.
My own tooling kept flagging untouched high-priority tasks as a failure; it had the facts right and the judgment exactly wrong.
After months building autonomous agents, the model choice rarely mattered as much as the memory system did. Here is what actually carries the load.