Operator Notes.
How we run a company on agents — what works, what breaks, and the numbers behind it. No fluff.
What is agent looping — and why one pass is never enough
A single model call produces a draft. A loop produces a result. The difference is the checking step, and it changes what you can safely automate.
Who runs an agent company: the marketing engineer
The role that makes agent operations work is neither a marketer nor a developer. A walkthrough of a week in the role, with the real dashboards.
Open vs. closed agents: two modes, two contracts
Open agents explore with a human in the loop. Closed agents run unattended on a schedule and check their own work. Confusing the two is how automation projects die.
The company brain: what to ingest first
Not everything your company knows belongs in agent memory. Start with the four document classes agents actually read, and skip the data lake.
AI slop is a systems problem — the eval loop
Generic output is not a model limitation. It is what happens when nothing between the model and the reader checks the work. Build the checker.
Adding an agent to any vertical: the 4-week pattern
Sales, support, finance, logistics — the deployment pattern is the same four weeks. An annotated build log, from workflow map to first unattended run.
Operator Notes.
How we run a company on agents. One e-mail when we publish — no drip sequence, no sales follow-up.