My AI-assisted dev workflow, mid-2026

People keep asking what my “AI setup” looks like. Here’s the honest version as of mid-2026 — not the demo, the daily driver.

The shape of it

1. Agentic coding up front. I plan in phases, hand each phase to an agent, and review the diff before moving on. The leverage isn’t typing speed — it’s that I can hold a whole architecture in my head while something else does the careful, boring file-by-file work.
2. Self-hosted open models for volume. Anything high-frequency and low-stakes — summarization, classification, first-draft copy — runs on open-weight models I host myself. Cost goes to near zero and nothing leaves my infrastructure.
3. A spec the agent can’t drift from. Every project gets a short rules file: locked stack, no new dependencies without asking, smallest correct change.

The rule that makes it work

Plan in dependency order. Review every phase. Never let the tool widen scope on its own.

Agents are fast and confident, which is exactly why you keep them on rails. The phases force a checkpoint; the rules file forces restraint. That combination is what turns “impressive demo” into “shipped product.”

More on the self-hosting setup in a future post — it deserves its own writeup.