Lore

Unhinged Mode: Sandboxed Agent Autonomy

Overview

"Unhinged mode" describes letting an agent operate with maximal autonomy — minimal steering, no approval gates — inside a sandboxed Playground Codebase that only touches data and prompts, with no bash or system access. Because the blast radius of a wrong move is contained, the agent can be given far more freedom to try things than would be safe on a real machine.

Mechanism

Autonomy risk is treated as a property of the environment, not just the agent: the same agent that would be dangerous with unrestricted system access becomes safe to run "unhinged" once it is boxed into a playground where mistakes only cost a re-run, not a broken production system.

Apply

Before deciding how much autonomy to grant an agent, first build (or find) an environment where a wrong move genuinely can't damage real systems — an eval playground, a sandboxed data copy — then grant maximal freedom only inside that boundary.

Related: Eval-Building Workflow (Dataset → Prompt → MCP → Scorer), The Agent Line