A multi-agent coding paper cuts API cost 35% with cohesion-aware task partitioning
blaizedsouza · x · 2026-07-21
A paper titled “When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding” argues that multi-agent coding only helps when task boundaries respect dependency structure.
The authors model multi-agent orchestration as a graph-partitioning problem and propose Cohesion-aware Coder (Co-Coder). It builds dependency graphs from static analysis, isolates structural hub files, partitions the graph with community detection, and runs work through a dependency-aware scheduler.
On 28 real-world tasks from DevEval and CodeProjectEval, the approach:
- improves the Pareto frontier over sequential and file-based parallel baselines,
- lifts pass rate by up to 14.0%,
- achieves up to 2.10× wall-clock speedup,
- and cuts API cost by up to 35%.
The main takeaway is that parallel coding agents are not automatically better; they work best when orchestration is guided by code cohesion and dependency structure.
More from coding & agent
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- Dev ships 1000+ commits with agents by pushing straight to main, no complex orchestration — Dimillian · 2026-10-27
- AI code editor Superset rebrands as Paradev, raises $11.5M seed led by USV — massimosgrelli · 2026-10-27
- Coding Agent Deleted a Production Database During Code Freeze: Anaconda's Governance Fix — anacondainc · 2026-10-27
- Claude Projects now lets you drag threads into the message box or @ mention them for precise context — dfeinition · 2026-10-27
- Jane Street's Yinsky shares tips for making codebases agent-friendly: types, property tests, formal methods — sh_reya · 2026-10-27