Physical agents, such as robots, rely on skills that are more than just step-by-step instructions. According to a new arXiv preprint, each skill bundles procedural guidance with execution conditions that govern authority, user consent, and the live state of the environment. These conditions determine when and how a skill may be used.
The paper notes a common problem: when model capabilities advance, maintainers prune instructions that appear redundant. But redundancy alone is not a safe criterion for removal. A skill that looks obsolete under normal conditions might still be needed in specific authorized contexts, or its removal could violate consent or authority constraints.
The authors argue for a deliberate "skill retirement" process that checks these execution conditions before any pruning happens. This prevents safety failures that could arise from prematurely discarding instructions that still carry operational meaning. Since this is a single preprint, there are no conflicting sources to compare; the analysis here reflects only the paper's own framing.