Parallel agents collapse: Opus joint success drops from 62% to 16%
EliasEskin · x · 2026-11-04
The authors compare a single-agent oracle, sequential hand-offs, and parallel agents (unaware / aware / message-connected) across open and closed-weight models. Parallel agents fail far more: Opus 4.8 drops from 62% to 16% joint success on contradictory tasks and 86% to 57% on complementary ones (open-weight average 60% → 19%, 72% → 53%). Even with gold-test debugging, oracle hits 98% vs 38% for parallel agents — each half works, the combination doesn't. A single A→B hand-off recovers much of the gap (16% → 54%) but doesn't scale to hundreds of agents.
Related event: ParallACT benchmark: parallel coding agents drop to 16% success(9 posts)→
More from coding & agent
- Open-source 'Palantir for Cybersecurity' maps live exploits and routes fixes to coding agents — Fowe · 2026-11-04
- Matt Shumer opens his agent for SMS trials, giving $150 in free credits to new users — mattshumer_ · 2026-11-04
- Indie dev adds full MCP support to SnapWidget, from signup to payments — hermanschutte · 2026-11-04
- Forrester Predicts 2027: Unreviewed AI Code Will 'Blow Up' a Company, 15% of Developers Job-Hunting — rseroter · 2026-11-04
- Model Vendors Post-Train for Tool Search, Squeezing Out MCP-Layer Workarounds — pvncher · 2026-11-04
- Engineer builds agent that clears years-old code backlog without any mandate — BradPorter_ · 2026-11-04