Last weekend, I built a rough kanban board for agentic projects.
Nothing fancy. Basic board for tracking features and tasks. But the goal isn't the board itself. It's to connect an AI agent to it so the agent can work through the backlog autonomously and develop the system itself.
For the last few years, once a year, I've tried to build a fully automatic software pipeline — one that doesn't need a real developer. It's never worked.
Until now, I think it might.
The goal: an agent picks tasks from the backlog, writes code, tests it, verifies it works, and commits if the tests pass. No human in the loop. Just autonomous development.
It's crude. It's slow. It makes mistakes. But it's functional. And with each iteration, it gets a little better.
Fifteen months ago, I generated 10 websites for $2 and asked LinkedIn which one was best. Today, I'm building systems where AI agents pick their own work and execute autonomously.
The distance between those points is enormous. But looking back, the path was surprisingly straight. From AI as tool, to AI as pair programmer, to AI as junior developer, to AI as team member, to AI as autonomous agent.
Each step required more trust and less control. Each step was terrifying the first time and obvious in retrospect.
Have I coded myself out of a job? Maybe. But I think I've coded myself into the next one.
Part of the #100DaysToOffload documenting agentic development in 2026