Cohesive Systems

Interesting paper: coding agents do well on tasks they can script their way through (grep, AST rewrites). Accuracy falls when they have to reason about behavior, like identifying which lines in two nearly identical classes actually behave differently. On SWE-bench, the number of reading and analysis steps correlated more strongly with failure than lines edited. Finding code isn’t the same as understanding its behavior. Better comprehension tools will help…but why not make system meaning explicit so agents don’t have to reconstruct it from implementations? arxiv.org/abs/2610.10610

The AI era has given formal methods a reemerging role. As code production becomes increasingly automated, formal methods are moving upstream, from verifying code to formalizing intent. The scarce resource is no longer writing software; it’s expressing precisely what the software should do. That makes executable specifications, constraints, and semantic models far more valuable than they were when humans wrote every line of code.