A one-person company where every other seat is an agent.
I run a one-person product company. Every other seat is an agent. I set the architecture, write the briefs, design the tests and own every release. My own agent has worked for me every day for over a year; since July 2026 it does so under an authorization and audit layer I designed and built.
A worked example: this site
This site went from a blank page to what you are reading in one day. One person, eight agent sessions. This is how the work was split.
- 1 day
- blank page to this
- 8
- agent sessions, one person
- 40
- personal sites studied first
- 21
- product screens rendered by the real app
Research, in parallel
Four research sessions ran at once. Between them they opened forty personal sites in a real browser and read how hiring managers actually go through a portfolio. Each report had to separate what it saw from what it inferred.
Build, with the real product on screen
One session built the site. Another drove the real Soul Engine app with demo data and rendered twenty-one screens from it. Nothing on the product pages is a mock-up.
The calls stay with me
Which three products lead, what the first screen says, what gets cut: those were my calls. The agents made it fast. They did not decide.
The rules I run it by
Many sessions, each with a brief
A coordinating session breaks the work down. Research sessions, build sessions and review sessions run in parallel, each with a written brief: what to do, what done looks like, what not to touch, how to report. Two sessions never write the same files.
Memory that survives the context window
Every project carries a small set of files — map, status, decisions, handoff — so a fresh session starts where the last one stopped. I open-sourced the method as project-brain.
Reports are claims until checked
An agent’s report has to separate what it verified from what it read somewhere. I spot-check the load-bearing facts before anything goes out.
Gates on anything that cannot be undone
Agents act freely on what is reversible. Everything else stops at a gate I built: each grant needs my Touch ID, works once, and is pinned to one action on one target, so approving A never allows B. Agents use API keys without ever seeing them. Every grant, use and refusal is recorded and tamper-evident, and 375 automated tests hold those boundaries. While I am at work the agent runs its own shift and splits the work among sub-agents; payments and signatures always come back to me.
I still make every call
What to build, what to cut, what trade to accept, when it is good enough to ship. The agents make me faster. They do not make the decisions.