Technology

Plurai aims to improve AI agent performance with a method borrowed from self-driving car development

Benchmarks suggest AI agents still need close supervision for complex work, with top models falling short of full task accuracy in LiveBench 2.0 and an EON corporate simulation.

Sylvester Eyamah

By Sylvester Eyamah

Nov 14, 2026, 7:31 PM  ·  1 min read

Plurai aims to improve AI agent performance with a method borrowed from self-driving car development
Plurai aims to improve AI agent performance with a method borrowed from self-driving car development — Technology
Although AI agents have become a popular focus, their real-world abilities often fall short of what their makers suggest. Tests designed to measure how well agents handle complicated assignments indicate they generally cannot be trusted to operate with minimal oversight. According to LiveBench 2.0, Claude Fable 5.2 succeeded on 82.9 out of 100 tasks, while ChatGPT-6 Astra succeeded on 75.6 out of 100. Separately, AI infrastructure firm EON evaluated agents in a simulated business setting; among the models tested, Claude Opus 4.8 achieved the highest task accuracy, at 76.8%.
Newsletter

Get the biggest stories in your inbox

A concise morning briefing, curated by Chronicle editors.