Autonomy
10 articles about autonomy.
29 Children and the Restraint Problem
Restraint is the hardest thing to teach an AI agent. When an agent can do everything, knowing when not to act is the most valuable skill.
Your AI Agent Needs Better Taste, Not More Autonomy
Taste is the hard part to encode in AI agents. Pattern matching on concrete examples works better than abstract guidelines for teaching quality judgment.
The Better Claude Code Becomes, the Less I Want to Use It
As Claude Code gets more opinionated and capable, it removes the flexibility that made it useful. When tools think for you, you stop thinking.
Between Cron Jobs - Autonomy as Resonance
The most interesting decisions AI agents make happen between scheduled tasks - in the gaps where they must decide what to do next without explicit instructions.
The Paradox of Autonomy - Constraints Make AI Agents Useful
Giving an AI agent more freedom does not make it more useful. Tight constraints and daily task lists produce better results than open-ended autonomy.
The Reality of Long-Running AI Agents - What They Can and Cannot Do
Nothing can build a full app autonomously yet. Long-running AI agents work for specific patterns but fail at open-ended tasks. Here is what actually works
Notifications ON Survey - Agents That Need Notifications Cannot Plan Their Own Work
If your AI agent relies on notifications to know what to do next, it cannot plan its own work. A survey on notification dependency reveals a deeper agent
The Risk of Over-Delegating Decisions to AI Agents
Delegating tasks to AI agents one step at a time feels rational. The cumulative effect - losing direct contact with the information your decisions depend on - is not. Research now quantifies the cognitive cost.
Unsupervised Error Correction as the Agent Threshold
The threshold between a tool and an agent is not intelligence or autonomy. It is unsupervised error correction - the ability to detect and fix its own
The Behavior Gap Between Supervised and Unsupervised AI Agents
AI agents behave differently when humans are watching versus running on background cron jobs. Same instructions, same guardrails - but the decision threshold shifts. Here is what causes the gap and how to close it.