Running agents in parallel with branches
Set up a machine once, snapshot it, and branch it into independent copies so several agents can attack the same task side by side.
Dock team · · 4 min read
Running several agents on the same problem is one of the most effective ways to get a good result: try three approaches, keep the best. The hard part is giving each agent an identical starting point without paying for setup three times.
Set up once
Start one dock and prepare it: clone the repository, install dependencies, seed the database, warm the build cache. Then archive it. The disk is saved as a snapshot, and that snapshot is the starting point for everything that follows.
Branch per attempt
Branching creates a new dock from a snapshot. Each branch gets its own ID and is fully writable, and changes in one branch never affect the source or any sibling. Give each agent its own branch and let them work in parallel.
Because every branch records the snapshot it came from, you always know which attempt started where. That lineage matters when you compare results or need to explain how a change was produced.
Keep the winner, archive the rest
When the attempts finish, keep working in the branch that produced the best result and archive or delete the others. Archived docks have no compute charge, so losing attempts cost only the time they ran.
The same pattern works for evaluations: branch one snapshot per model or prompt, run them side by side, and compare from identical starting states.