Decision
A weekly “agent systems manager” cron (Saturdays 9am ET) evaluates all agentic systems on evidence (job health, output quality, value delivered, coverage gaps) and DMs Michael ranked improvement proposals. Propose-approve model: the manager never self-modifies prompts, schedules, skills, or scripts; Michael approves proposals by number and changes are implemented with him in session.
Context
Michael’s directive: the brain strategy must make the company self-improve through intelligent systems, like a manager for the agent workforce. Runs Saturday so approved changes can land before the Sunday scoreboard/QA cycle.
Options considered
- Self-modifying agent (auto-applies improvements): rejected for now; approval gate keeps a human on the quality loop and prevents drift. Revisit only with a strong track record of accepted proposals.
Consequences
- New log: Agent Systems Manager Log (findings, proposals, Michael’s verdicts). Rejected proposals are logged so they are never re-proposed.
- The manager’s scorecard becomes the “are the systems working” section of record.