Human-agent collaboration in Buzz: missions, not constant orchestration

Human-agent collaboration works when the human supplies direction and authority, while the system owns decomposition, state tracking, dependency checks, proof collection, and recovery.

If the owner has to repeatedly ask whether documents were read, workers exist, tasks started, or proof was collected, the operating loop is incomplete.

The owner should choose business direction, approve major gates, review results, and decide ambiguous product or commercial questions.

The owner should not have to act as scheduler, dispatcher, proof inspector, duplicate suppressor, dependency tracker, or reminder system.

Agents should read the mission, inspect current state, find blockers, locate independent work, dispatch only needed workers, capture proof, ask for inspection, and queue the next safe job.

When an intervention is required, the system should record why the human had to intervene and convert that failure into a permanent check or regression test.

Channel messages should contain work that changes state: assignment, proof, blocker, decision request, verdict, deployment, or next job. Bare acknowledgements add noise and can retrigger workers.

Mentions should be used for action, not narration. The right message reaches the accountable person and includes enough evidence to avoid a follow-up interrogation.