Agent Autonomy

Autonomy Is a Spectrum, Not a Switch

Agent autonomy is the range of control an AI agent holds over its own next action — running from fully scripted (a developer decided every step in advance) to fully autonomous (the agent decides what to do and when it's done, with no per-step check). It's a continuous line, not a fixed set of tiers with hard walls between them.

Five Points Worth Naming

• Scripted automation — zero freedom; a fixed code path runs the same way every time
• Suggest-only — the agent proposes, a human decides; nothing changes until it's accepted
• Approve-each-step — the agent plans, but stops before every action for a yes or no
• Approve-checkpoints — the agent runs a batch of steps on its own, pausing at review points instead of every action
• Fully autonomous — the agent acts, observes, and decides its next move without stopping, reporting back once done or stuck

Cursor's Read-only, Ask, Plan, and Agent modes, and Claude Code's Manual, Accept Edits, Plan, Auto, and Bypass permission tiers are real, working instances of this same dial.

Picking the Right Level

No point on the dial is inherently better. The right level depends on the stakes of the task: how bad it would be if the agent got it wrong, how easy that mistake is to undo, and how certain you already are about the right outcome. High-stakes, hard-to-reverse, uncertain tasks need tighter approval; low-stakes, reversible, well-understood ones can run with more freedom.

Autonomy can also be granted by accident, not just on purpose — a misread setting or a bug can silently widen an agent's permissions beyond what a user believes they've approved.