Research Roundup: Multi-Agent Coding Makes Results Worse — Use Single Agents for Write Tasks

PilgrimofHaqq2 · reddit · 2026-09-05

A systematic roundup of research on parallel vs. sequential agent systems: parallel multi-agent wins for read-heavy, independently sliceable work (research, search, review), while sequential single agents are better for single-chain decision tasks like coding and writing. Key evidence:

For parallel

Anthropic's multi-agent research system beat a single agent by 90.2% on internal research evals, at 15x token cost; Anthropic itself notes coding has fewer parallelizable tasks.
LangChain: read tasks parallelize, write tasks shouldn't.

Against multi-agent for coding

Nature Machine Intelligence (Jul 2026): across 260 controlled configs, every multi-agent variant made coding worse (-1.3% to -12.8% on SWE-bench Verified); above a 45% single-agent baseline, gains go zero-to-negative; error amplification hit 17.2x without centralized verification.
UC Berkeley MAST (NeurIPS 2025): 41–86.7% failure rates across 7 popular multi-agent frameworks (1,642 real traces) — design/coordination faults, not model limits.
Cognition: split-context parallel workers make conflicting implicit decisions; Devin uses one single-threaded agent plus context compression.
arXiv (Apr 2026): under equal thinking-token budgets, single agents match or beat multi-agent teams.
Princeton "AI Agents That Matter" (TMLR): complex setups cost up to 100x more than simple baselines that perform equally well.

Middle ground: OpenHands — one strong generalist agent covers most use cases; go multi-agent only when you need isolation, and keep teams small.

Original post →

Research Roundup: Multi-Agent Coding Makes Results Worse — Use Single Agents for Write Tasks

More from coding & agent