Control Design

An agent's autonomy isn't a switch — it's a dial. Approval gates, autonomy levels, escalation paths, kill switches: the control design that makes autonomy shippable.

▶ Watch this reel

What you'll learn

Approval gates
Autonomy levels
Escalation paths
Kill switch, rollback & audit

Remember this

Approval gates trigger on consequences and parameters, render what-not-which previews, batch intelligently, and stay rare enough to be read
Autonomy is a per-dimension dial — suggest → recommend → act → act-and-report — earned through measured reliability and demoted on regression
Escalations carry complete packages and feed the learning loop; kill switches are independent, drilled, and paired with rollback-by-design

Approval gates

Consequence + parameter aware · show WHAT (rendered preview/diff) not which tool · batch as consolidated diffs · rare enough to be read.

Autonomy levels

Suggest → recommend → act (bounded) → act-and-report.
Per-dimension, per-user; earned via measured success, auto-demoted on regression; versioned config artifact.

Escalation paths

Triggers: repeated failure, low confidence, novelty, policy, user demand.
Package: tried/outcomes, state, assessment, artifact refs.
Escalations → golden-set cases; rate = health metric.

Kill switch & rollback

Kill: seconds-to-zero, infra-independent, manual, drilled.
Rollback: reversible-by-design or gated; runbook pre-written; post-incident audit from existing trails.

Code: The control plane as configuration

domains:
  customer_refunds:
    autonomy: act                      # earned: 99% success over 500 runs
    bounds: { max_amount_usd: 200, reversible_only: true }
    gates: [ { trigger: "amount > 200", action: require_approval },
             { trigger: "recipient_count > 10", action: require_approval } ]
    escalation: { on: [failures>=3, confidence<0.6], package: full }

  infrastructure_changes:
    autonomy: recommend                # not yet earned
    gates: [ { trigger: "always", action: require_approval } ]

global:
  kill_switch: { scope: [per_agent, all], drill: quarterly }
  rollback: { backups: automatic, runbook: v3, signoff: oncall_lead }
  audit: { append_only: true, includes: [gates, levels, escalations, actions] }