KedgeFlow ties agent actions to exact permissions with single-use claims

KedgeFlow places an independent controller between an agent’s intent and target effects. Instead of asking the model whether a tool call is allowed, an issuer signs exact parameters and a shared ledger consumes a grant only once. Testing confirms useful concurrency properties within the chosen mock-target setup. An overly permissive policy still permitted an undesirable operation, so the design does not magically discover correct user intent; it enforces previously chosen authority.

A one-use grant can prevent the same authorization from being consumed twice, but cannot ensure an external service did not commit before a timeout. A valid signature also proves conformity to the issued policy, not that the policy captured the user’s intent.