One agent is enough.
The first product avoids company administration and agent fleets. Each private classroom owns one clear agent and one history.
Teaching an agent should create inspectable state, not invisible memory or authority nobody can review.
The first product avoids company administration and agent fleets. Each private classroom owns one clear agent and one history.
AGENTS.md, knowledge, teacher explanations, procedures, corrections and YAML rules remain visible—not hidden conversational state.
A proposed training file is not active. A draft is not published. The interface always shows the real status.
After you press Process, the agent asks about authority, scope and exceptions instead of inventing certainty. Binary questions get quick replies.
Current public demonstrations are entirely fictional. Private uploads, emails and output files do not appear there.
Publishing, sending, buying and changing external systems require a separate human-controlled execution step.
The framework creates durable agent files, source-grounded context, revisions and approval history. A self-hosted installation may connect a model, but the files remain the authoritative training record.