Agents can implement, review, and deploy software. Responsibility still has to live somewhere.
You can read this post here for free.
In one of my recent experiments, I let a coding agent develop software across several work cycles while I was on a long drive. I prepared the instructions, boundaries, and checkpoints before leaving. The agent handled the implementation, and I reviewed its progress while parked every time I stoped for gas or stretch my legs.
In another project, I stayed much closer to the work. I was working on an evaluation system for a healthcare recommendation platform, manually reviewing architectural decisions and implementation details with the help of ask mode in my IDE.
Both are part of how I now build software. After testing these approaches for a while, I have a better sense of when to use my AI-first framework, when to give an IDE agent a task, and when to work through a decision myself with AI assistance. The choice depends on the task and the setting I’m working in.
My AI-first framework organizes development around agent roles and explicit artifacts. Requirements become a specification, architectural work produces a design, and implementation proceeds through defined handoffs. This works well in a team where responsibilities and decision authority need to be explicit. My Obsidian agent workflow keeps plans, constraints, and decisions in Markdown that I can inspect and carry between tools and sessions…