This week's briefing

What happened this week

This week, AI’s progress moved beyond model behavior into the systems around it: agent memory, browser-native tools, inference hardware, and clinical use. The defining pattern was controlled deployment, with validation gates, experimental standards, and single-case evidence setting the terms that carry into next week.

Lead story / benchmark

Google’s WikiSkill lifts agent scores without retraining models

Google Research introduced WikiSkill, a framework intended to help AI agents retain practical lessons from earlier attempts without retraining the underlying model. In tests across five benchmarks, Gemini 3.5 Flash’s average score rose from 49.5% to 68.1% when it used the system’s evolving skills.

WikiSkill separates evidence from action. A Raw Layer keeps full execution traces, a persistent Wiki Layer turns those records into documented failures and successful strategies, and a Skill Layer holds the instructions used on new tasks. Proposed changes are tested on a separate validation set; if an update reduces performance, the system can restore the prior skill while retaining the experience behind the failed proposal.

The strongest results came in math and spreadsheet tasks. Gemini 3.5 Flash rose from 33.0% to 72.6% on LiveMath and from 50.5% to 76.6% on SpreadSheet, while Qwen-3.6-27B’s five-benchmark average increased from 39.4% to 63.3%. Improvements were smaller on the OfficeQA document question-answering task.

The result offers a route to agent improvement based on preserving operational evidence and testing better instructions, rather than changing model weights. But it is not a universal reliability fix: larger models generally benefited more, smaller ones could revert to default behavior on long multi-step searches, and cross-model skill transfer was inconsistent. Deployment still requires testing whether a learned skill carries to the model and task at hand.