LLM & Agents · 2026
primo-turno: Keeping AI Coding Agents on Track
AI coding agents are incredibly productive. Until they wander off. So I built a small supervisor that keeps them honest.
→A Go supervisor that runs coding agents one verified feature at a time, until the spec is done.
iteration 07 → pick feature implement · validate ✓ · commit fresh context → iteration 08
Highlights
- ✦One change per iteration, validated & committed
- ✦Progress log that each new iteration reads first
- ✦Supervisor, not agent, decides when the job is done
- ✦Built spec-first, with numbered tickets
Why I built this
People ask me how I ship so fast.
Part of the answer: I let AI agents do a lot of the typing. But never unsupervised.
How it works
It’s based on the “Ralph loop” pattern:
- Give the agent a spec
- It picks the one most important next feature
- It implements, validates and commits
- Fresh context, repeat until done
primo-turno is the supervisor around that loop, written in Go. It owns the control flow. The agent never decides on its own that the job is finished.
The lesson
Agentic coding isn’t “let the AI do everything”.
It’s small steps, clear specs, real tests. Same as good engineering always was. Just a lot faster.
That’s the engine room behind every client project I ship.