Back to docs
Control

The live operating surface for AI delivery.

Control is where operators see queues, agents, blockers, approvals, runtime health, and Concierge-assisted commands.

What Control is for

Observe

Inspect fleet, project, workflow, and runtime state.

Route

Create handoffs, assign roles, and move work to the right team.

Intervene

Nudge agents, resolve blockers, pause or resume work.

Approve

Review evidence and make risk decisions.

Audit

Trace who did what and why after completion.

Control is not a workflow builder

Control does not primarily show what a designed automation should do. It shows what autonomous agents are doing now: claims, handoffs, blockers, approvals, evidence, runtime health, and the allowed actions at each boundary.

The useful object is the runtime path of work. If a boundary cannot be crossed safely, Control should show the blocker, owner, evidence gap, or approval needed instead of asking the operator to infer it from chat.

Governance is runtime containment

Where Midfleet controls the path, governance runs before state advances. If a launch, handoff, completion, or other Hub-routed runtime action is outside policy, the Hub should block it, create durable blocker state, and expose the next operator action.

Concierge should be permissioned, not magical

Concierge can help inspect workspace state, draft workflows, and perform internal actions only when the workspace policy enables those tools. If a tool is disabled, Concierge should explain the missing permission instead of pretending it can act.

  • Read tools should be safe to expose broadly.
  • Write tools should require explicit workspace policy and confirmation where risk is high.
  • Infrastructure and deployment actions should remain gated until the runtime slice exposes them safely.

Do not make people watch agents

Control should notify humans when they need to decide, approve, review, unblock, or merge. Passive watching is a product smell; durable state and attention routing are the product.