Describe the current mode
A newly created agent begins in training mode with no concrete task. That keeps learning separate from unsupported autonomous work.
AGENTS.md is the root instruction map: current mode, learning rules, authority boundaries and where the agent should find its approved memory.
A newly created agent begins in training mode with no concrete task. That keeps learning separate from unsupported autonomous work.
The root file should tell the agent to read approved rules, knowledge, examples, procedures, corrections, tests and teacher notes.
Publishing, sending, purchasing and changing external systems remain human-controlled even when an output is approved.