Core concept · supervision

Human authority

Human supervisors grant operating authority. Agents propose work, execute supported actions, preserve exact states, and return exceptions.

Runtime authorization

Enforced controls

Agent credentials cannot exercise human approval routes. Cancelling a schedule and pausing activity are the safe direction: a workspace key may call POST /v1/posts/:id/cancel and POST /v1/pause, and the audit log records actor_type: api_key. Publishing, approving, rescheduling, and resuming stay human-gated.

Human supervisor

Approval and policy

Approve plans and posts
Set trust and account policy
Pause and resume activity
Resolve recoverable exceptions

AI agent

Bounded operation

Propose plans
Submit supported work
Cancel scheduled work
Pause activity
Start connection handoffs
Read states and measurements

Denied to agent

Authority changes

Approve itself
Expand trust
Resume paused activity
Publish now or reschedule
Exercise human approval routes

Connection ownership

Intended boundary

Human-owned connection

Humans connect destination accounts from the authenticated Accounts screen when connection is enabled for their workspace.

The agent may read workspace status and destination IDs, then create an authorization URL for the human. Workspace-key creation, social authorization, accepted media delivery, and MCP process setup remain human-owned.

Conceptual policy

Authority example

This example explains the authority boundary. It is not a callable schema or MCP configuration.

CONCEPTUAL BOUNDARY · NOT CALLABLE

enforced:
  agent:
    - propose_plans
    - submit_supported_work
    - read_states
    - read_measurements
    - start_connection_handoffs
    - cancel_scheduled_work
    - pause_activity
  human:
    - approve_plans
    - set_trust_and_policy
    - pause_activity
    - resume_activity
    - resolve_exceptions
  denied_to_agent:
    - approve
    - expand_trust
    - resume
    - publish_now
    - reschedule

intended_boundary:
  account_connection_owner: human
  customer_connection_ui: authenticated
  workspace_enablement: required
  agent_connection_tool: connect_account
  agent_connection_action: authorization_url_only

Supervisor return

Required handoffs

requires_approval

Approval decision

Return the exact state and resource ID to the human.

draft · missed_approval · missed_schedule

Schedule recovery

These post states require a human publish or reschedule decision.

plan · changes_requested

Plan resubmission

Use returned feedback for a deliberate new plan submission.

post · changes_requested

Post handoff

Return the exact state and feedback to a human. No configured post-revision call exists.

failed · partial · rejected

Outcome review

Return exact post and destination evidence. Do not recast the result as success.