AI Progress Measured by Agent Autonomy Time

Dan Shipper's "leash length" theory provides a framework for understanding AI advancement and predicting when we'll reach AGI by measuring how long AI can work autonomously before requiring human intervention.

The Evolution of AI Autonomy

• AI progress can be measured by how long you can "give it a leash" to work independently

  • Early AI (like Copilot): Tab completion - immediate supervision required
  • ChatGPT: Single question/response cycles - minimal autonomy
  • Current agents (Claude Opus 4, Gemini): Can work for 20-30 minutes without intervention
  • Future AGI: Will work indefinitely without supervision

• The progression mirrors human development from infancy to adulthood

  • Infants: Completely dependent on caregivers
  • Toddlers: Can be left alone briefly but need frequent supervision
  • Children: Gradually gain independence for longer periods
  • Adults: Function autonomously for indefinite periods

Defining AGI Through Economic Value

• AGI will be reached when it becomes "economically profitable for people to run agents indefinitely"

  • The agent never turns off
  • It's always working on something valuable
  • It responds when needed but doesn't require constant direction
  • The value it creates exceeds the cost of running it continuously

• This definition ties AGI to practical utility rather than abstract capabilities

  • Not just about passing tests or mimicking human abilities
  • About creating enough economic value to justify permanent operation
  • Requires both technical capability and economic efficiency

Implications for AI Development

• Current AI is at the "toddler stage" of development (20-30 minutes of autonomy)
• Each generation of AI extends this leash length
• The progression isn't just about raw intelligence but about reliability and trustworthiness
• Companies building AI should focus on extending autonomous operation time
• Users should track how long they can trust their AI tools to work independently

Practical Applications Today

• Organizations at the cutting edge are already using multiple agents working in parallel
• Teams are developing systems where agents can work for extended periods
• The most advanced workflows involve agents that can:
  • Create their own to-do lists
  • Take notes on their progress
  • Process large amounts of information autonomously
  • Judge their own work quality and improve it before returning results