Agents

Autonomy Budgets

Share

Autonomy budgets is an AI interface design pattern that grants unattended agent execution with hard bounds on time or action count, such as 'run for 30 minutes or 50 actions, whichever comes first, then pause for review.' This UX pattern prevents runaway sessions, the failure mode where an agent loops on a misread goal and accumulates changes or costs far beyond the user's intent, by forcing a natural checkpoint the user has to approve before more autonomy is extended. Budgets are visible during the run, decrement legibly as the agent works, and default to pause rather than continue when exhausted. The pattern pairs naturally with reversibility marking and cost meters: it gives the user a bounded window to trade for speed without signing a blank check on either time or impact.

Use Case

Essential for long-running AI coding agents, research agents, and autonomous workflows where bounded unattended execution prevents runaway loops and keeps humans in the loop at natural checkpoints.

Examples in Wild

Cursor Agent max iterationsClaude Code session capsDevin work unitsGitHub Actions job timeouts

Use this pattern in your project

Copy this prompt to generate a production-ready implementation in Cursor, Claude Code, Lovable, or any AI coding agent.

Generate a production-ready implementation of the "Autonomy Budgets" AI interface design pattern.

Pattern Description:
Interactive Demo
Restart demo
Autonomy Budgets
Time budget6 / 30 min
Action budget12 / 50 steps

We’ll stop the run when either limit is reached.

Get new patterns by email

Weekly AI interface UX notes and resources on Substack, no spam, unsubscribe anytime.

Subscribe on Substack