AI agents that scrape economic data from the web usually start from zero every time, even on sites they have visited before.
Researchers built EconSkills, a system that turns verified successful browsing sessions on a benchmark called EconWebArena into reusable 'skills' - step-by-step procedures with placeholders swapped in for site-specific values, plus built-in checks for verifying results and recovering from dead ends. The team tested two separate questions: does a skill that clearly matches a new task actually help, and can an agent still find and use the right skill once it has to pick one from a larger library instead of being handed it directly. When agents got a matching skill directly, they succeeded more often and in fewer steps than agents working from scratch, and turning a raw trajectory into an abstracted procedure worked much better than replaying the old steps verbatim. Once agents had to retrieve a skill themselves from a growing library, though, performance slid back down to roughly match the no-skill baseline, with gains showing up mainly on tasks the library already covered directly.
That gap is the real finding here. It is not that procedural memory fails - it clearly helps when matched correctly - it is that matching itself is the bottleneck, and approximate matches on unfamiliar tasks quietly cancel out the benefit seen on covered ones. Any agent that repeatedly hits the same class of sites, whether for stock data, government filings, or shopping carts, will run into the same wall as its library of known moves grows.
It is a familiar problem dressed up in new language: reusable macros have always worked great until something has to decide, on the fly, which macro actually applies.