Admission Control for Agent Skills and MCP Supply Chains

Skills, MCP servers, and agent runtimes are emerging as behavior-bearing software supply-chain releases rather than trusted listings. The actionable pattern is immutable release identity, capability-based review, static plus semantic plus sandboxed dynamic testing, synthetic credentials, provenance, continuous revalidation, and expiring risk acceptance. This connects tool governance with memory poisoning defenses and external policy enforcement, but standard manifests and interoperable policy tooling remain open questions.