Google and Mercor’s PLE Exposes Constraint Bottlenecks in Engineering Agents

What happened

Google and Mercor’s Physical-Engineering's Last Exam (PLE) evaluates 85 engineering tasks and finds a gap between local accuracy and viable design delivery. Claude Opus 5.5 with Claude Code scored 48.3% on 56 hidden tests; across co-creation tasks, it met 91% of individual constraints on average, while only 41.3% of design submissions met all hard safety and feasibility requirements. The article examines constraint coupling, design state management, and verification as challenges for engineering agents.

Read next

• More changes in today’s briefingOpenAI Releases Images 2.6 With SVG Export, GIFs, and Faster Generation

The story so far

Google and Mercor introduced Physical-Engineering's Last Exam (PLE), a benchmark of 85 engineering tasks for evaluating AI agents. On 56 hidden tests, Claude Opus 5.5 with Claude Code scored 48.3%. Across co-creation tasks, it satisfied 91% of individual constraints on average, but only 41.3% of design submissions met every hard safety and feasibility requirement. The results point to a gap between meeting constraints one at a time and delivering a viable design that satisfies them all.