Agent Skills
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Agent Skills are modular instruction sets (typically stored as .md files) that guide AI agents through specific workflows or tasks, enabling reusable, discoverable capabilities that can be selectively invoked during agent execution. They are a distinct abstraction from scripts or tools, designed to provide contextual procedural knowledge to a model at runtime.
What Skills Are — and Aren't
Philipp Schmid (Google DeepMind) draws a clear boundary between skills and scripts: if a workflow is fully deterministic and always executes the same way, a script is more appropriate than a skill. Using a skill in that scenario wastes model tokens unnecessarily. source
Skills are distinguished by their use of progressive disclosure — they surface relevant procedural detail only when needed, rather than loading all context upfront. source
Performance Impact
According to Skills Bench results cited by Schmid, skills on average improve agent performance by roughly 15%. source This positions skills as a meaningful but not transformative lever — one that depends heavily on implementation quality.
Structural Constraints
Skill .md files should be kept below 500 lines to avoid performance degradation. source Beyond this length, the signal-to-noise ratio degrades and the model's ability to correctly interpret and apply the skill diminishes.
The Triggering Problem
Schmid identifies incorrect triggering as the dominant failure mode: 50% of skill failures occur because the skill was not invoked at all, due to a weak or ambiguous skill description. source
The remedy is writing directive descriptions rather than passive informational ones — the description must tell the model explicitly when and how to use the skill, not merely describe what it does. source
The Eval Gap
A central critique running through Schmid's talk is the near-total absence of evals in the existing skills ecosystem. Of the 50,000+ skills indexed by Skills Bench, almost none had accompanying evals; most were AI-written and effectively untested. source This is framed as a systemic quality problem — shipping skills without evals means operators have no validated signal that skills work as intended.
Points of Disagreement
No competing speakers or contrary positions appear in the current graph material. Schmid's framing is consistently cautionary: skills offer real gains but are routinely undermined by poor description design, excessive length, and the absence of evaluation infrastructure.