From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution

The paper's core provocation is that richer documentation actually hurts reusability: smaller, more abstract experience representations survive structural code changes better than verbose skill packages, which is the opposite of how most agent memory systems are currently designed.

This sits in direct tension with the direction Google and OpenAI are both taking in production. Google's Skills feature in Chrome, covered here two days ago, bets on prompt templates as the reusable unit of AI behavior. OpenAI's upgraded Codex, also from this week, leans into persistent memory and expanded context as the substrate for agentic continuity. Both approaches resemble what this paper calls the 'Skill' paradigm, the one that underperformed. The MIT Technology Review piece on enterprise AI as an operating layer is the better conceptual neighbor here: it argues that the real competition is over the infrastructure where AI is refined over time, and Gene-style representations are essentially a proposal for what that refinement substrate should look like at the model level.

Watch whether any of the major coding agent teams (OpenAI Codex, Anthropic Claude Code) publish ablations comparing compact versus documentation-heavy memory formats in the next two quarters. If they do and compact representations win there too, this paper's framing will have moved from academic to directly actionable for product decisions.