The finding I keep coming back to in Google’s new paper is not the headline number. It is this one. Skills evolved by one model work when you run them on a different model, from a different family. Sometimes better than the skills that model evolved for itself.

Read that as a procurement fact and it gets interesting fast.

If your operational knowledge lives in the model weights, switching vendors is a migration. If it lives in files the agent wrote, switching vendors is a config change.

What Google actually measured

WikiSkill came out of Google Research and Virginia Tech on August 27. It splits three things that normally get mashed together: raw execution traces, the knowledge distilled from them, and the executable skills an agent runs.

The loop is unglamorous. The agent runs tasks. Successes and failures both get mined for patterns, and those patterns go into a persistent wiki. Skills get rewritten from the wiki, and a rewrite that makes things worse gets rolled back. The next round starts from the accumulated record instead of from nothing.

Across five benchmarks that have very little in common (LiveMath, SealQA, SpreadSheet, OfficeQA and ALFWorld):

• Qwen-3.6-27B goes from 39.4% to 63.3%, a gain of nearly 24 points.
• Gemini-3.5-Flash goes from 49.5% to 68.1%.
• Gemini on LiveMath alone climbs from 33.0% to 72.6%, and on SpreadSheet from 50.5% to 76.6%.

It also beats the prior skill-evolution methods it was compared against, namely Trace2Skill, EvoSkill and SkillOpt.

Then the headline. Qwen-3.5-9B with evolved skills averages 47.4%, above the 39.4% of a Qwen-3.6-27B running without them.

Fine print worth stating, because the paper states it. Larger models gain more from the method than smaller ones. Scale still buys you something. The claim is not that size stopped mattering. It is that accumulated knowledge competes with buying more size, at a fraction of the cost.

The ablation is the real result

Strip the wiki out and keep the skill files, and performance drops hard.

That single experiment relocates the value. It is not the skills. It is the persistent record they are derived from, which means a company that copies the skill files without building the record gets very little.

Andrej Karpathy argued for exactly this shape months ago, with no measurement behind it. We wrote about it at the time in knowledge compounds or rots. The argument has numbers now.

What the enterprise version looks like

Not a research artifact. Something much duller.

A place where failure patterns get written down. The real exceptions in your operation, the ones no vendor documentation contains. The cases your agent handled badly and what the correction was. Plain text, versioned, readable by a human who was not there. We covered how to curate that material in the wiki for agents, and we described how ours works in the IQ Source brain.

Ours predates this paper, which means for a year I was defending an intuition. Now I can point at a table instead. That is a better position to argue from and I will take it.

Here is the part that should decide your quarter. You will change models at least twice more before this year ends, and each change resets whatever the model itself had absorbed. The record does not reset. Start it in September and by December it is worth something; start it in December and you are where you are today.

Let’s build the wiki your agents write to