Isolated worktrees
A worktree is a separate checkout of the same repository. Each agent ran in its own worktree and its own branch, meaning its changes never collided in real time.
MoreProofCase Study: Agent Orchestration
A five-agent website redesign without overlapping file ownership.
In an earlier redesign of this site, I split seven pages across five coding agents. Separate checkouts and a shared brief let me combine their changes without merge conflicts.
Built with
Git worktrees.
System: brief-governed parallel execution
01
The coordination problem
Parallel agents overwrite each other
When multiple agents edit the same repository, they create conflicting diffs, overwrite shared dependencies, and invent contradictory design rules.
In an earlier redesign session, seven pages needed an update. Executing it serially would take too long, but simply launching multiple coding agents in the same checkout would corrupt the repository. The challenge was organizing the work so that independent agents could execute in parallel without stepping on each other's files.
02
Partitioning
Who owned which files
File-disjoint partitioning
Agents were assigned specific pages. Shared global files were frozen and could not be edited by the agents.
The coherence contract
A shared design brief served as the single source of truth, establishing rules before any agent started working.
03
Human integration
How integration was checked
The agents wrote the code on their own branches, but a person reviewed and merged it.
Each agent self-verified with lint and typecheck before committing to its branch. I then merged all five branches—with zero conflicts due to the disjoint file assignments—and ran the single production build. I calibrated every page from screenshots in both themes and ran an accessibility sweep before making one reviewable pull request.
04
Outcomes
What the run demonstrated and did not demonstrate
This sprint is an account of a historical session. It demonstrated that strict file ownership and separate checkouts can make parallel agent work coherent. It is not proof that all agent-written code is correct by default, nor does it imply that the current site has no later human edits. Subsequent updates continue to be shaped by human judgment.