After this chapter you will be able to define a trust progression with guardrails and human checkpoints that determines when an agent can run unsupervised. Deploying agents safely is fundamentally a problem of earned autonomy: an agent should not receive more freedom than its demonstrated reliability justifies, and the transition from supervised to unsupervised operation must be deliberate rather than accidental. The trust progression A trust progression is an explicit ladder of operating modes, each with a defined scope of action. A typical ladder begins with the agent proposing actions for human approval, moves to the agent executing low-risk actions autonomously while flagging anything consequential, and ends with the agent operating unsupervised within a bounded domain. Each rung is earned, not assumed: an agent advances only after accumulating evidence — a measured error rate on a representative workload, clean audits of its past decisions, and stable behavior under edge cases. Crucially, the progression also runs in reverse. When an agent violates a guardrail or produces degraded output, it is demoted to a more supervised mode until confidence is rebuilt. Guardrails as the boundary of autonomy Guardrails translate trust into enforceable limits. Structural guardrails constrain what the agent can touch: scoped credentials, sandboxed execution, allowlists of tools and data sources, and hard ceilings on spending, volume, or blast radius. Behavioral guardrails constrain what the agent can do: policy checks on outputs, classifiers for prohibited content or actions, and deterministic refusal rules for irreversible operations such as deletions, payments, or external communications. The design principle is that unsupervised operation should be safe by construction — a failure at the least-trusted level should be annoying, not catastrophic. Human checkpoints Human checkpoints are the scheduled moments where oversight is guaranteed rather than optional. They include pre-approval gates for irreversible or high-consequence actions, periodic review of sampled agent decisions, and escalation triggers that page a human when uncertainty, cost, or anomaly thresholds are crossed. Checkpoints should be placed where errors are most expensive and hardest to undo, and each checkpoint should generate a record — an audit trail of what the agent did, why, and what the human decided — that feeds back into the trust evaluation. Deciding when an agent can run unsupervised The decision to grant unsupervised operation is a conjunction, not a single metric. Define it in advance: the agent must exceed a target success rate on a benchmark that matches production distribution, operate for a soak period without guardrail violations, and the domain itself must have bounded, reversible, low-consequence actions. Write the criteria down, make promotion an auditable decision, and keep a kill switch that reverts to supervised mode instantly. With a documented progression, enforced guardrails, and meaningful human checkpoints, autonomy becomes a managed outcome of evidence rather than a leap of faith.