Skill Retirement
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Skill Retirement is the practice of removing a previously implemented agentic skill or capability from a system once an underlying AI model has improved to the point where that skill is no longer necessary for reliable performance.
Core concept
Philipp Schmid (Google DeepMind) advocates for Skill Retirement as a natural lifecycle stage in agentic system development, tightly coupled to evals. The premise is that many skills are scaffolded workarounds for model limitations — once a model is capable enough to handle a task natively, the compensating skill becomes redundant overhead. According to Schmid, evals are the mechanism that makes this determination rigorous rather than speculative: "Evals will tell us when we can retire skill and when not."
Evals as a safety net for retirement decisions
A key nuance in Schmid's position is that retiring a skill does not mean discarding its associated evals. He advocates retaining evals even after the skill itself is removed, so that future model regressions can be detected promptly. If a model update degrades the capability that justified retirement, the eval will surface the regression and the skill can be reintroduced: "you can reintroduce the skill". This makes Skill Retirement a reversible, evidence-driven decision rather than a permanent one.
Relationship to the broader evals philosophy
Skill Retirement is one concrete application of Schmid's broader thesis — encapsulated in the talk title Don't Ship Skills Without Evals — that every skill in an agentic pipeline should be justified and governed by measurable evaluation criteria at every stage of its lifecycle: creation, operation, and retirement.