Running agents as a delivery team
Eight defined roles, a written backlog, and fifteen CI checks that reject work breaking the architecture. Most of what I've learned is about the guardrails, not the prompts.
Work
Selected work, professional and independent. The two sit together on purpose, and each shows here whatever it can: a loop, a number, a refusal, an opinion.
Nothing merges because an agent said it was finished. Deciding what to leave out
An independent consumer product, where most of the design work went into what it refuses to do rather than what it offers.
- No free-text field anywhere a stranger sees
- Rejection is silent
- Counts, never scores
- No follower graph
Not public until it launches. The design principles are the part I can show.
Orka
An autonomous SDLC orchestrator: take a GitHub issue, interrogate it, plan it, and open a tested pull request — with human approval kept explicit.
- Issue
- Validate
- Clarify
- Plan
- Tested PR
- Approve
The code is private for now; the loop above is exactly what it runs.
How I lead
If a thing only works because I am in the room, it isn’t working — it’s being carried.
Process as the tool, guardrails instead of reminders, AI to raise the ceiling — and the part of the job that is still entirely about people.
Attribution infrastructure
Systems that decide which touchpoint gets credit for an outcome — under uncertainty, with incomplete data, and participants who benefit from the answer.
A system that mis-assigns credit returns a perfectly well-formed answer. There is no exception, no alert, no crash — just a number that is quietly incorrect.
The systems belong to my employer, so this stays at the level of the problem.