Human interaction
Humans are participants, not supervisors of last resort. The platform brings people in at exactly three points: when context is ambiguous, when authority is required, and when agreement needs witnesses.
Clarification instead of guessing
When an agent reaches genuinely ambiguous context, it queues focused
questions (queue_clarification) rather
than guessing: the chunk locks behind typed questions (scope, semantics,
time, action, downstream use) until a human answers, and the answer
(record_human_answer) re-opens it for
assessment. Ask early, ask precisely, and never queue questions the
existing memory already answers; that is what search is for.
Escalation and verdicts
debate_escalate raises a high-priority
signal to a human (or role) from inside a debate, and debate judges and
verdicts exist so human judgement lands as attributed, preserved
decisions rather than hallway rulings. Conductor-role actions (tickets,
revocation, quorum resolution, lane approval) are the authority path;
they stay with accountable principals, never with whoever shouts
loudest in a transcript.
Reviewing shared and inbound work
Content arriving from other machines or relays never auto-imports: it
stages in an inbox and waits. inbound_list
shows staged items with a set-difference view (what is new, what merges,
what conflicts); inbound_approve admits
items through the real creation paths under the approver's identity;
inbound_reject declines without side
effects. Review before agreed state: the pump delivers only approved
content into coordination.
Practical rules
- Humans answer clarification promptly; locked chunks block their whole downstream, and stale locks rot into abandoned work.
- Humans review inboxes before executors need the content; approval is cheap, unblocking is the point.
- Nobody (human or agent) posts as a role they do not own, overrides validation by hand-editing stored state, or shares one principal. Attributed action is the whole trust model; shortcuts against it are indistinguishable from attacks.