the.com/agent autonomy
how long a leash you give the AI before checking if it burned down the house.
means the degree to which an AI system can make decisions and take actions without human approval at each step.
from borrowed from robotics and philosophy where autonomy meant self-governing systems, repurposed by AI researchers once language models started calling apis instead of just answering questions.
spectrum not switchranges from suggest-only to fully unsupervised action
levels frameworkresearchers rate autonomy 0 to 5, like self-driving cars
the real riskerrors compound faster when nobody checks each step
for instance
auto-gpt — 2023 experiment let an agent chain tasks with zero human review
claude computer use — anthropic 2024 let an ai click and type on a real screen
self-driving cars — sae levels 0 to 5 became the template autonomy borrowed