ParallACT benchmark: multi-principal agent coordination stumps frontier models at 16% success
EliasEskin · x · 2026-11-04
Researchers at UT Austin (with Peter Stone and Joydeep Biswas) introduce ParallACT, a benchmark for multi-principal multi-agent coordination: what happens when many developers each delegate their own tasks to agents in large monorepos with 1000+ contributors?
Unlike mainstream multi-agent setups that decompose one centralized request, ParallACT's requests come from different principals and can genuinely conflict.
Key findings:
- Frontier models struggle badly: Opus 4.8 achieves only 16% success, even when agents can talk to each other
- In 15% of cases, one agent did both agents' work while the other stepped aside
The authors note this works fine for two agents but becomes a serious problem at scale with hundreds of agents.
Related event: UT Austin's ParallACT Benchmark: Parallel Agents' Success Crashes to 16%(9 posts)→
More from coding & agent
- Zero-shot mobile manipulation demo stacks Dream-1 policy, MCP planner and ROS nav — scott_e_reed · 2026-11-04
- Anthropic launches Claude auto-connect beta, letting admins onboard whole orgs to MCP plugins — TansuYegen · 2026-11-04
- Coding agents build baselines for 36 top-journal AI methods, matching ~95% of scores — owl_posting · 2026-11-04
- Postman desktop auto-installs Codex/Claude plugins, tries to trust its own startup hook — jedisct1 · 2026-11-04
- Oracle's agentic AI product lead: enterprise AI roadmaps run on outdated use-case lists — akumar · 2026-11-04
- Sentdex: vibe coding is real, OpenHands shows dev may not involve reading code — Sentdex · 2026-11-04