The Road to Agent Autonomy: Challenges, Discoveries, and a 28-Line Solution

Some CLAUDE.md or AGENTS.md files I've seen are micromanagement dressed up as engineering. Hundreds of lines dictating which files to read, which commands to run, what order to do things in. Some hit 1000s of tokens before the agent starts thinking about your actual task. Every token spent on instructions is a token the agent can't spend on your work. And the more rigid the instructions, the more brittle the agent becomes. I developed a 28-line operating law for autonomous agents — born from wanting to give the agent a task in Claude Code terminal and have it complete it without any input or feedback, no matter how easy or complex. Here's what I discovered. The plan mode trap I asked the agent to build a full-stack app. Set it loose with --dangerously-skip-permissions and waited. 34 minutes later: zero files created. The agent had entered plan mode — voluntarily switching itself to read-only. It spawned three research subagents, produced excellent analysis, then sat there. Planning. Un