Codeuteronomy
Large work is divided and fenced
Parallel agents multiply throughput only when the boundaries are explicit. Without fences they collide, duplicate, and overwrite.
Two things break on large tasks. The work outgrows a single context window: early decisions get forgotten and quality decays late in a long run. And the obvious fix, several agents at once, adds a new failure wherever their territories touch: two subagents editing one tree undo each other's work, each blind to the other, and you find out in a tangled diff.
One move fixes both. Split the task into pieces one agent does well, and give each piece a boundary it may not cross: this directory, this file, not the others. Stated plainly, parallel subagents compose; left implicit, they compete. On big work your job is partitioning it and guarding the seams.
How to practice it
Spawn a subagent per piece
(the Task/Agent tool), each with its fence written into the prompt: “you own X; do not touch Y.”
Give concurrent writers their own git worktree
so parallel edits merge instead of clobbering one another.
Size each piece to one context window.
If the agent would forget its own start before finishing, split further.
Define recurring roles under .claude/agents/
so a fenced specialist is one spawn away.
Review the seams first.
That is where the integration bugs live.