Do skills teach procedures or inject missing facts?

The paper asks "when do skills help, why do they work, and where do they fail," and answers that skills mostly work by anchoring procedure, not by supplying knowledge. Its abstract puts it directly: "skills work when noisy trajectories become procedural anchors that stabilize execution." Procedural anchoring "accounts for 65.7% of skill cases, versus 4.5% for explicit knowledge injection, showing that skills stabilize action rather than inject missing facts." It also reports that skills improve over Workflow Memory by 6.06 points in matched comparisons. The conclusion pairs this with the failure conditions: skills fail "when they are retrieved incorrectly, invoked in the wrong context, followed too rigidly, or used on tasks that require deeper reformulation and runtime validation."

The evidence is a contrastive design. Controlled experiments isolate "representation, outcome annotation, retrieval difficulty, and crossframework robustness," and paired trajectory analysis then explains why a skill helped or hurt in a given trial. The authors normalize 8,135 trial records, keep 238 valid unique labels from 240 open-coded records, and consolidate them into "a taxonomy of three high-level categories and twelve skill-use modes." The introduction supplies the reasoning behind the anchoring reading: repeated failures in tool-using agents "often arise not from a lack of high-level reasoning, but from rediscovering the same procedural details again and again," and a skill is "a compact description of what to do, what to check, and what pitfalls to avoid." If the missing ingredient is procedure rather than fact, a skill acts as a fixed reference that a noisy trajectory can be held against. The same property explains the failures. An anchor helps only when it is the right one, in the right context, and loosely enough held that the agent can still reformulate.

This gives a mechanism to notes that so far report gains without isolating their source. Can agents learn reusable sub-task routines from past experience? credits reusable sub-task routines for the improvement, and the anchoring finding fits that reading while adding a head-to-head comparison against "Workflow Memory," though the excerpt does not say which implementation. Can frozen language models continually improve through memory structure alone? argues that the shape of textual memory matters more than generic hints, and the 4.5% share for knowledge injection points the same way. Can a separate trained curator improve skill libraries better than frozen agents? observes that verbose generic guidance gave way to actionable structures as curation improved, which is the same distinction seen from the curator's side. The lifecycle framing also qualifies designs that treat skill generation as the bottleneck, such as How can agent systems share learned skills across users?. The paper's warning is that "building better self-evolving agents requires not only generating more skills, but also improving how agents represent, retrieve, and leverage procedural knowledge."

The excerpt is silent on several things that bear on how far the claim travels. It does not name the benchmarks, harnesses or LLMs, does not define what counts as a "skill case" in the 65.7% and 4.5% shares, lists only two of the twelve skill-use modes, and gives no success-rate effects beyond the single 6.06-point figure. It also cuts off before the three advantages skills are said to promise. What it supports is narrower and still useful: evaluate skills by what they do to trajectories, not only by aggregate success, and treat retrieval, invocation context and how strictly a skill is followed as design decisions in their own right.