Agent patterns

Tool-Call Policy Enforcement Agent

Governance evaluated after the fact cannot prevent the action it is meant to govern — it has to sit in the path of the call itself.

Human review point

Only for actions a policy resolves to 'flag for review' — blocked and allowed actions proceed without waiting on a person, by design.

Must not do

Allow an action a matching policy resolves to block, regardless of which other policy might separately allow it — most-restrictive-wins is not optional.

Inputs

Attempted tool call; calling agent identity; configured Agent Policy set

Output artefact

An Agent Access Finding recording the outcome, logged before the call is permitted or denied