Do AI agents actually learn from experience?

Yes, but not in the way the question suggests: the underlying model stays identical, no weights are retrained. What changes is the set of procedures the agent carries with it, and skill evolution is the process that produces them. The agent rereads its own execution traces, the successful attempts alongside the failed ones, extracts the regularity that explains them and writes it down as a skill: an executable, inspectable artifact rather than a memory. On the next similar occasion that skill gets retrieved and applied, so the improvement is cumulative and visible in a file instead of buried in parameters. The idea originates with Voyager, which in 2023 gave an agent a skill library of code built during exploration; the 2026 research line takes it out of simulated environments and into real work with tools, documents and business systems.

Procedural memory, not declarative memory

This is the boundary that makes skill evolution a subject of its own. Memory consolidation works on facts: it dedupes memories, resolves contradictions, decides whether to overwrite or version a value. Here the output is not a fact but a procedure, something that runs and can fail at runtime. In the taxonomy of memory types it is the procedural branch, the one dispatched in a single line because until recently it had no mechanism of its own: now it does. There is also a difference of authorship against hand-packaged skills, where a team's procedural knowledge is written by a person. Here the agent writes it, and the two paths coexist in the same repository.

An enterprise example

An agent preparing month-end accounting reconciliations keeps getting foreign suppliers wrong: it applies the exchange rate of the posting date instead of the document date. During review the error is corrected by hand, and the correction lands in a trace like any other. Skill evolution promotes it to a procedure: which field to read, in what order, which check to run before closing. The following month the agent does not repeat the mistake because it follows its own skill, not because somebody reminded it in the prompt. The economic value sits in the next step. In this work, evolved skills transfer across models and across model families, and a small model equipped with them can outperform a much larger model without them: that is the thesis of WikiSkill at Google, with SkillOpt at Microsoft Research reporting the same transfer across models and execution harnesses. The practical implication, which the paper does not measure, is that you compile the library once with the expensive model and serve production with the cheap one.

Why it matters for decision-makers

Transfer is not free, and it should not be bought as a slogan: it is uneven, because general procedures hold up while workarounds tuned to the source model can hurt whoever inherits them, so a library compiled elsewhere is a starting point, not a finished asset. The deeper limit is more serious, and the entry on self-play already states it: without an external oracle verifying the outcome, self-improvement risks reinforcing nothing but the model's confidence in its own answers. A skill born from a trace nobody validated encodes a mistake and makes it repeatable. So the question to ask whoever supplies the system is not whether the agent learns, but who signs off on promoting a trace to a skill and where that is written down.