ParallACT benchmark: multi-principal agent coordination, Opus at 16% success
EliasEskin · x · 2026-11-04
Most multi-agent benchmarks split one global request into compatible subtasks, but in real large repos 1000+ devs each raise independent, sometimes conflicting requests their agents don't even know about; routing everything to a central agent is a bottleneck and single point of failure. The authors introduce ParallACT: starting from real SWE-Bench Pro tasks (Task A) paired with a separately-motivated Task B touching the same code, each agent sees only its own task. Two task types: complementary (must agree on a shared interface to pass integration tests) and contradictory (incompatible edits to the same code, but a joint solution exists). Frontier models struggle — 16% success for Opus 4.8, even when they can talk.
Related event: UT Austin's ParallACT Benchmark: Parallel Agents' Success Crashes to 16%(9 posts)→
More from coding & agent
- GhostDraw: a hobbyist turns hand-drawn sketches into app diagrams with tldraw and AI — Dhananjay_Tech · 2026-11-04
- CoworkOS Debuts: Three-Layer Agent Coworker Framework Lifts Qwen3.6 by 9.76 Points — KyeGomezB · 2026-11-04
- DOLORES: Training-Free Meta-Scaffold Lifts 8B Models Past 32B Baselines — StellaLisy · 2026-11-04
- Berkeley-affiliated StarTrail open-sources a deep research benchmark and paper-search toolkit — CShorten30 · 2026-11-04
- AI job-search framework built on Claude Code hits 69.3k GitHub stars — adnan_hashmi · 2026-11-04
- DuplexAgent-RSI separates live speech from async agent work in voice assistants — jiqizhixin · 2026-11-04