ParallACT Benchmark: Frontier Models Hit Just 16% on Multi-Principal Agent Tasks
EliasEskin · x · 2026-11-04
EliasEskin and PeterStoneTX introduce ParallACT, a benchmark of multi-principal multi-agent tasks inspired by large repos with 1000+ concurrent devs delegating to agents. Unlike typical multi-agent setups that decompose one centralized request, ParallACT tasks come from different principals and can conflict. Frontier models struggle — Opus 4.8 succeeds only 16% of the time, even when agents can communicate — highlighting coordination and conflict resolution as the key unsolved difficulty.
Related event: UT Austin's ParallACT Benchmark: Parallel Agents' Success Crashes to 16%(9 posts)→
More from coding & agent
- Compound Engineering 3.35 encodes repo learnings as tests and lint rules, not just docs — kieranklaassen · 2026-11-04
- A nightly wiki-builder agent: prompt included, built for biotech tacit knowledge — nlarusstone · 2026-11-04
- Understanding your code unlocks better AI prompts, especially for 0-to-1 builds — brandon_xyzw · 2026-11-04
- Terence Tao predicts an explosion of experimental mathematics as AI maps math difficulty — haider1 · 2026-11-04
- Put static secrets in your HTTP proxy: exe.dev on agent-era key management — davidcrawshaw · 2026-11-04
- Claude to auto-connect plugins that trust it as an issuer, cutting agent re-auth friction — testingcatalog · 2026-11-04