Five agents, one frozen file

During an earlier seven-page redesign session, five agents running in five git worktrees (separate checkouts of the repository) created five branches. At the end, an integrator merged all five into main with zero merge conflicts. Not few. Zero.

That number had nothing to do with model quality. Agents in the same repo collide for the same reasons two junior engineers on the same file collide, except agents type faster and never stop to ask. Everything that worked came from three rules written down before any agent started.

• Scopes were disjoint by file, not by topic. Each agent owned named files outright. “You handle the demo page, you handle the experience page” works; “you handle spacing, you handle color” guarantees two agents in one file by lunchtime.
• The one file everyone needed was frozen. Every page wanted to touch the global stylesheet, so nobody got it. Shared changes were specified in advance and landed once, by the integrator. A frozen file is an ugly rule that outperforms every clever alternative.
• Coherence came from a committed brief, not from chat. One document defined the three shared recipes and the hard bans, and every agent read the same one. Style drifts when five workers each hold a private copy of the plan in their heads.

Resume, don’t respawn

One more rule earned its keep the hard way. Mid-sprint, agents died against usage limits. The recovery that worked was resuming the same agent with its context intact rather than spawning a fresh one on the same task. A replacement agent re-derives everything, makes different micro-decisions, and hands you a subtle fork of your own conventions. Continuity beats a clean start almost every time.

None of this is exotic. It’s the same discipline you’d impose on any parallel work crew. The lesson from the sprint is that agents don’t exempt you from coordination. They just raise the price of skipping it.