Cheezy World

In The New Team I described what a single AI-driven product team looks like: one Product Owner and one or two Developers, working in tight feedback loops that produce shippable software in hours rather than weeks. That post was about the atomic unit. This post is about what happens when you zoom out.

Most product organizations were not designed around a single team. They were designed around the assumption that delivering software is slow, expensive, and risky. Every layer of the org chart - the directors, the program managers, the design ops leads, the release coordinators - exists to manage that slowness and risk. When the underlying work becomes fast and cheap, those layers do not automatically become useful. Many of them become friction.

As many of you know, I’ve built a tool called Stride. It is a human / AI collaboration tool that offers a robust workflow with many features to drive collaboration and predictability. It takes the form of a hosted kanban-like app that manages the detailed tasks and agent plugins for all of the major AI agents. Even though it has several thousand users now, I have struggled with forcing the agents to always follow the workflow defined in the plugins. The plugins define a lifecycle that is decomposed into seven skills — one orchestrator (stride-workflow) and six sub-skills that handle individual phases (stride-claiming-tasks, stride-completing-tasks, and so on). The sub-skills exist because each phase has a non-trivial API contract that the agent has to get exactly right. The plugins also contain four subagents and a hook designed to execute developer defined commands at specific times during the lifecycle of the work.

At the AI Coffee Chat this past Wednesday we were having a great conversation about how to minimize the overall context that your Agent uses and what happens as the context gets full. I will not go into the details here of what the Agent context is and why it is important to not fill it up with unnecessary things. In the Chat session, I made the statement that a great way to minimize context usage is to understand the core thing you are asking your Agent to do and try to move everything else out to SubAgents, Hooks, or Skills.