Scheduling planner improves integration of parallel coding agents in preprint

Researchers report that a planner built into Nerveplane increased clean integration from 1/9 to 9/9 scenarios and reduced merge conflicts from 13 to 0 in a benchmark of parallel coding agents. The preprint evaluates coordination through actual git merges, using deterministic simulations and live agents. The planner partitions declared work into separate scopes and orders merges so that producers finish before consumers. On a live breaking contract change, clean integration rose from 0 under both baselines to 1.0 on a frontier model and 0.6 on a small model. Routing facts to agents did not improve long-context accuracy at window-fitting scales.