# Nerveplane

The authors reframe coordination among parallel coding agents as a scheduling problem: before agents start, partition work into disjoint scopes from each item's declared scope and order merges along the producer-consumer dependency graph; they implement this planner in Nerveplane and evaluate it with NP-Bench, a three-arm benchmark grounded in a real git merge, where deterministic clean-integration success rises from 1/9 to 9/9 and merge conflicts fall from 13 to 0, and a live breaking contract change rises from 0.00 to 1.00 on a frontier model and 0.60 on a small one.

The authors reframe coordination among parallel coding agents as a scheduling problem: before agents start, partition work into disjoint scopes from each item's declared scope and order merges along the producer-consumer dependency graph; they implement this planner in Nerveplane and evaluate it with NP-Bench, a three-arm benchmark grounded in a real git merge, where deterministic clean-integration success rises from 1/9 to 9/9 and merge conflicts fall from 13 to 0, and a live breaking contract change rises from 0.00 to 1.00 on a frontier model and 0.60 on a small one.