The core shift was simple: stop asking one assistant to do everything. Instead, assign narrow intent and bounded tools per role. Writing, implementation, research, operations, and release validation became separate lanes.
That change immediately improved predictability. Fewer wrong-tool selections. Clearer style per output. Easier review. Safer rollback when a lane misbehaved.
The hidden win was observability: once each lane had a definition of done, diagnostics became objective instead of emotional. You can fix a lane; it is far harder to fix a vague personality drift.
Design for handoffs, not heroics. Specialist agents with explicit contracts produce calmer systems and cleaner outcomes.