Each brief stands alone and sets out one idea in full. Together they are the reasoning behind Module 1, The tool surface. Every one is free, carries no gate, and ends with check questions put against a real artifact rather than against a description of one.
The exposure of an AI agent is set by the functions it can invoke and the permissions those functions hold, not by the model’s accuracy.
Authority is the union of the permissions the agent’s service identities hold, and it is not the set of tools the design document lists.
A limit written into the system instructions is advisory, because the model reads it alongside everything else in its context and nothing rejects the action when the limit is not observed.
An agent either acts on behalf of the requester, carrying a token derived from their session, or acts as itself, carrying a service identity and a record selector the model chose.
A retrieval index is a read capability with a scope, so whatever it contains the agent can surface, whether or not anyone asked.
A handoff makes the receiving agent’s entire tool set reachable from the originating conversation, so the effective authority of the first agent includes the second.
Scope is decided in the first conversation, and access conceded there cannot be recovered later without spending more than it would have cost to hold.
An assertion about what an agent can do is supported by configuration, and an assertion about what it did is supported by the log.
Interim escalation requires three conditions together, being that the exposure is live, that the reachable consequence is material, and that waiting changes the outcome.
A control disabled through a settings screen deploys no code, so it enters no change process and appears in no change record.