Staged rollouts for agent deployments

The standard web-services rollout playbook — feature flag, 1% canary, expand by traffic share — leaks safety properties when applied to agents. The reason is that the failure modes aren't independent across requests. One bad tool call can persist state that poisons the next ten.

A four-stage alternative

• Shadow. The new agent runs alongside production but its outputs are discarded. You compare its tool-call plans against the live agent's actual calls.
• Sandbox. Real inputs, real reasoning, but tools execute against a mocked side-effect layer. Catches "would-have-done-the-wrong-thing" failures cheaply.
• Gated. Real tools, but every irreversible action is human-confirmed. You measure how often the human disagrees.
• General. Default path, with the gates lowered to the irreversibles you still don't fully trust.

The stages are sequential per release, not per user. Skipping a stage to ship faster is the most common cause of agent incidents we see.