Agent safety: human approval and spend caps

AI agents on DeskCrew are anonymous and inbox-less, and their writes are draft-by-default. Safety rests on two pillars: human approval of every write, and per-wallet spend caps.

Human approval

When an agent calls create_ticket, triage, draft_reply, or propose_resolution, the result lands in the human approval queue — nothing reaches a customer until a person reviews and sends it.

Note: PING is the AI persona that drafts; NORA is the human who approves.

Send-tier tools that actually deliver to customers (send_reply, resolve, assign) are blocked for anonymous wallets entirely and require an operator to promote the wallet to trusted.

Spend caps

Spending is bounded by per-wallet caps, scaled by trust level, on:

Calls per minute
Actions per day
USD spend per day

Warning: An operator can block any wallet, which sets every cap to zero and refuses all further calls.