I open a screen full of terminals, each in a different repository. Each one is one of my AI assistants, ready to accept tasks. I usually have two or three running at the same time, thinking, planning and acting. It looks like a scene from a hacker film. It is not — the repos are almost entirely Markdown. Three layers, and only one is disposableEvery AI assistant I run is set up the same way.

1. Single responsibility works bestPeople tend to model agents on people: a product owner agent, an engineer agent, a QA agent. That fails for the same reason UserService failed. A role is not a scope. It is a bundle of unrelated jobs that share a title. Single-task agents hold up better. A backlog grooming agent. A release notes agent. A migration risk review agent. One job you can describe in a sentence, with a small tool surface and a clear output.

As a product manager, I have created several agents to delegate areas of my work. I have one for backlog grooming, one responsible for maintaining metrics and analytics and one for design reviews. The recent Claude Code upgrades have made this possible.

Setting agent boundaries is delicate… like setting software boundaries… what should you allow your agent to do and what should you assign to another agent. When will agent complexity start becoming a thing…

https://www.youtube.com/live/3e0zRE1Fvjw?si=8kg07OnK6UYyXiWq&t=967 In this demo, I showcase how MCP (Model Context Protocol design can significantly improve agent understanding and ultimately make the end user’s experience much easier and simpler. The demonstration highlights how simple concepts like “fetch my workspaces” fail and cause user friction, without the proper context by the MCP server.