AI agents can now learn 'skills' from code and papers, but what makes one safe to trust and run?

This explores what turns a distilled skill (know-how packaged from code, papers, or a person's expertise) from a plausible-sounding note into something you can check and an agent can safely run. The corpus points to four ingredients: evidence that stays attached, a procedure rather than a fact sheet, a creation process that matches where the skill will be used, and a trail you can audit.

Verifiability starts with where the skill comes from. Code2Skill lifts skills straight from GitHub repositories, checks each one with a source-blind reconstruction test, and keeps the executable evidence attached. That produced over a million auditable skills, and they beat skills derived from agent trajectories or documents (42.90 to 47.90 macro-average across eight benchmarks) Can source code replace experience as skill raw material?. Compact, verified skills also pay off when nothing else changes. Added to a fixed GPT-3.5 agent, they improved four benchmarks by 9–134%, supplying operational knowledge that neither the model nor the planning harness had Can distilled skills close the gap in ML research agents?.

Readiness for execution is mostly about procedure, not facts. Across 8,135 trials, procedural anchoring explained 65.7% of skill successes and knowledge injection only 4.5%. Skills steady what an agent does rather than fill gaps in what it knows. They fail when they're retrieved for the wrong situation, invoked out of context, or followed too rigidly Do skills teach procedures or inject missing facts?. A separately trained curator learns the same lesson: it moves a skill library away from generic, wordy additions toward actionable execution logic and strategies that carry across tasks, and it works across different executor backbones Can a separate trained curator improve skill libraries better than frozen agents?.

Where a skill is written matters as much as what it says. MUSE-Autoskill has the agent create skills from inside its own reasoning loop, so each one is grounded in the exact task context, immediate feedback, and runtime validation. In-loop skills reached 87.94% task accuracy and transferred to other agents with minimal loss Does creating skills inside the agent loop eliminate mismatches?. That transfer is worth measuring rather than assuming. Work on agent harnesses found that performance can shift dramatically when the same harness runs under a different executor, and that downstream task scores don't reveal harness quality by themselves Can language models build and maintain their own agent harnesses?.

The last piece is being able to inspect, correct, and constrain a skill. COLLEAGUE.SKILL treats distilled expertise as versioned files that can be inspected, corrected, and rolled back, not hidden prompt state. It also keeps what a person knows separate from how they act, so each can be audited on its own Can person-grounded skills remain auditable without hidden prompt state?. Constraints need the same rigor. In one test, naming a prohibition was not enough to keep protected tests unmodified. The boundary had to specify the protected state and be paired with restricted tools Can explicit authorization boundaries prevent agents from modifying protected tests?. Where running code isn't practical, structured reasoning can check code equivalence without executing it. It reached 93% accuracy on real agent code, though only for certain task classes Can structured reasoning replace code execution for RL rewards?.