ParallACT benchmark: parallel agents drop to 16% success on conflicting tasks vs 62% for one agent
EliasEskin · x · 2026-11-04
UT Austin (with NVIDIA) introduces ParallACT, a multi-principal multi-agent benchmark built from real SWE-Bench Pro tasks paired with independently-motivated tasks touching the same code. Key results:
- Parallel agents fail where a single agent succeeds: on contradictory tasks, an agent given both specs achieves 62% vs 16% for two parallel agents (open-weight average 60% → 19%); complementary tasks drop 86% → 57% for Opus 4.8.
- Even with gold-test debugging, the oracle hits 98% but parallel agents only 38% — each half works, the combination doesn't.
- Communication helps little and costs much: the best message channel reaches 32% at 2.23× oracle tokens; merely knowing a peer exists moves 16% → 26%.
- Agents discuss the right code (88%) and often agree (59%), but only 22% of messages are read before editing starts, and 61% of runs with agreement still end in merge conflicts.
- Sequential A→B hand-off recovers much of the gap (16% → 54%) but doesn't scale to hundreds of agents.
The bottleneck is knowing when to listen and how to follow through, not what to say.
Related event: UT Austin's ParallACT Benchmark: Parallel Agents' Success Crashes to 16%(9 posts)→
More from coding & agent
- Aohan: inference should be fast enough that the environment becomes the bottleneck — _arohan_ · 2026-11-04
- Agent Buddy: a desk gadget that shows your AI coding agents' status at a glance — DanWahlin · 2026-11-04
- CircleCI on parallel coding agents: not everything should run in parallel — rseroter · 2026-11-04
- YC-backed Simulithic hires founding engineer: one command writes all your e2e tests — KlausCodes · 2026-11-04
- Developer uses Grok Bot as his always-on domain broker — Baconbrix · 2026-11-04
- VerseZero launches a game platform that simulates worlds for humans and AI agents — hwchung27 · 2026-11-04