Refactored my workflow again.
The old Pi workflow was heavily coupled — tons of hooks and scripts trying to enforce a fixed dev SOP. One wrong step and everything would just collapse.Models have gotten strong enough that I’m comfortable leaning into them now.
New flow is fully decoupled. Just a few subagents handling context isolation and model routing, with agents.md + append_system.md steering the main agent’s behavior.
Simple tasks? Main agent just does them directly. Complex ones get researcher and explorer subagents called first (serial or parallel) to gather web + local info, then planner subagent builds the detailed execution plan.
