When agents talk to agents, people can easily lose sight of what's happening. Oversight must be designed in.
Risks
- Actions taken several delegation steps away from the user.
- Errors compounding as results pass between agents.
- Unclear accountability when things go wrong.
Approval Points
Require human confirmation before consequential actions: payments, bookings, contract commitments, messages to third parties. Remote agents can use input-required states to ask for confirmation relayed to the user.
Visibility
Show users which agents were involved, what each did and what data was shared. Traces spanning agents help operators diagnose problems.
Limits
Set spending caps, allowed actions and time limits for delegated tasks.
Accountability
Record which agent, operated by which organisation, acted on whose behalf. Contracts between organisations should cover agent behaviour.
Escalation
Agents should hand over to people when uncertain or when tasks exceed their authority.
Review
Periodically audit delegation logs for unexpected patterns.