Google WikiSkill lets 9B agents beat larger models on tasks

WikiSkill gets at a core bottleneck in the AGI race: how to let agents truly learn from experience without endlessly re‑training massive models. By treating execution traces as raw material, curating them into a wiki and then evolving explicit skills that can be invoked later, the framework turns a static model into something closer to a continuously improving worker. XenoSpectrum highlights that a mid‑sized Qwen‑3.5‑9B model, when armed with these skills, can outperform a much larger 27B counterpart on a suite of math, web, spreadsheet, office and interactive tasks. That is a powerful argument that we may not need ever‑larger models to get better agents; we need better ways to store, edit and test procedural knowledge. ([xenospectrum.com](https://xenospectrum.com/en/google-wikiskill-agent-memory/))

Strategically, this shifts some of the innovation frontier from raw training runs to tooling and evaluation. If frameworks like WikiSkill generalize, smaller labs and application developers could squeeze much more value from commodity‑sized models by layering structured skill evolution on top. That is both an opportunity and a challenge for frontier labs: it makes their raw models more valuable when paired with good tooling, but it also gives competitors a lever to close the gap without matching their compute spend. Over time, we should expect the “model vs skills” split to mirror operating systems and apps: foundation models become platforms, and most differentiation comes from the skills and workflows layered above.