Autonomy Slider

Karpathy introduced the current LLM-product framing in his June 2025 Software Is Changing (Again) talk. He illustrated partial autonomy with Cursor's progression from completion and bounded edits to agent mode, Perplexity's search depths and Tesla's automation levels. The underlying idea is older: human-factors research has long modeled automation as a continuum across information acquisition, analysis, decision selection and action, while adjustable-autonomy research studies transfer of control between people and agents.

Sources: s1, s2, s3, s4, s5

The pattern gives product teams a vocabulary for progressive delegation. Low-autonomy interaction can make output easy to inspect; higher-autonomy modes can absorb longer workflows when their error cost is tolerable. The important design question is not simply how much the agent can do, but which decisions and actions remain with the person. Research on adjustable autonomy also shows that asking for human input has costs: delays, interruption and coordination failures can matter alongside the cost of an incorrect autonomous action.

Sources: s2, s4, s5

A coding tool might offer completion, a reviewable single-file edit, and a repository agent. Moving upward should not silently grant production credentials or remove review. The team can keep the agent in a sandbox, limit writable paths, require approval for network or deployment actions, run tests, show diffs and preserve rollback. Autonomy should rise only after task-specific evaluation demonstrates acceptable behavior; a user preference or mode name is not evidence that the model is competent for a consequential task.

Sources: s1, s3, s6

Autonomy is multidimensional: an agent can plan broadly yet lack write access, act repeatedly while requiring selected approvals, or run for a long time inside a narrow sandbox. Compressing those differences into one level can hide consequential permissions. Users may over-trust a high-autonomy label, approve prompts mechanically or lack enough context to verify output. Conversely, excessive intervention can stall coordination. The slider therefore does not create graceful fallback, calibrated trust or safety by itself. Consequential domains require explicit policy, least privilege, monitoring, rollback and human accountability beyond the interface mode.

Sources: s3, s4, s5