Agents need business meaning in a form they can use without improvising.
A human analyst can ask around when a metric is ambiguous. An agent often fills the gap with whatever context is closest: a stale doc, an old dashboard label, a copied formula, a confident guess. That is how automation turns semantic mess into faster semantic mess.
The semantic layer should expose compact, current context for agents. Not every document. Not a giant wiki dump. The agent needs the definition, owner, source, freshness rule, permission boundary, and allowed use for the concept involved in the task.
If an agent is drafting a renewal summary, it should know which account object matters, which health score is approved, where contract value lives, what counts as open risk, and whether internal risk labels can be mentioned externally. If it is investigating product usage, it should know which events are canonical and which have known instrumentation breaks.
This is not the same as the AI context layer, though the two touch. The semantic layer owns the business meaning. The context layer decides how that meaning is packaged, permissioned, refreshed, and supplied to agents for a given action.
The practical requirement is simple: definitions must be machine-readable enough to constrain behavior. A paragraph in a slide deck is better than nothing, but it is not enough if agents are expected to act safely.
Agent readiness is a useful forcing function. If the business cannot explain a term clearly enough for an agent to use, people are probably making shaky assumptions too.
This is part 9 of 10 in The Semantic Layer.
