Agentic systems are gaining a new architectural layer. An agent's capability no longer has to live inside a prompt, memory, or harness. MASkills shows how a skill can be made a separate versioned artifact with measured utility, validation, and rollback. If a change degrades quality, it never reaches the active library.
But a managed skill does not yet guarantee trust in the result. This is where ClaimReceipt comes in. The work separates the presence of logs from the presence of sufficient evidence. For a specific claim you have to check whether the stored data is enough and whether it covers the whole declared process. If the evidence is incomplete, the system should not make a confident conclusion.
These two ideas add up to a full lifecycle: a capability is created, used, measured, changed, passes validation, gains evidence, and only then qualifies for production.
This looks less like a "learning agent" and more like an Agent Capability Supply Chain. Capability becomes versioned, verifiable, and governed almost the same way as a software component. And this is exactly where Skill-as-Code, Evidence-as-Code, Evals-as-Code, and Architecture-as-Code begin to converge.