The token cost scaling point is real. Ran a 3-agent team last week - orchestrator plus two specialists - and watched the cost jump fast once they started sharing context. Switched to tighter scoping per agent and it helped. The parallel code review tip (security / performance / test coverage split) is something I hadn't tried. Going to test it this week.
One thing I found: the instruction file contradicting itself is a bigger failure mode than agent errors. Agents follow instructions too literally - so inconsistencies compound.
Quick question: when you say you use tmux rather than iTerm2, do you mean you rely on tmux for pane/window management instead of iTerm features? I assume you’re still running tmux inside iTerm anyway?
the tip about keeping each agent scoped to one responsibility is the one that took me longest to learn. kept trying to build god-agents that could do everything and they just got confused after 3-4 steps
The token cost scaling point is real. Ran a 3-agent team last week - orchestrator plus two specialists - and watched the cost jump fast once they started sharing context. Switched to tighter scoping per agent and it helped. The parallel code review tip (security / performance / test coverage split) is something I hadn't tried. Going to test it this week.
One thing I found: the instruction file contradicting itself is a bigger failure mode than agent errors. Agents follow instructions too literally - so inconsistencies compound.
Thanks for the great post!
Quick question: when you say you use tmux rather than iTerm2, do you mean you rely on tmux for pane/window management instead of iTerm features? I assume you’re still running tmux inside iTerm anyway?
the tip about keeping each agent scoped to one responsibility is the one that took me longest to learn. kept trying to build god-agents that could do everything and they just got confused after 3-4 steps