AI Operator Agents: CEO, Dev & Coach (human-in-the-loop)

🤖 Agents that operate a real business

These are not chatbots: they are fixed-role agents that run several times a day on a production SaaS, perceive the state of the business, propose actions, and execute only what the owner approves.

🎯 The fleet

• CEO agent: analyzes business metrics (opportunities, revenue, churn risk per salesperson), maintains a goal roadmap, and proposes strategy.
• Dev agent: takes direction from the CEO, proposes concrete technical improvements, and implements them after approval — deploy and verification included.
• Personal Coach: reads each salesperson's CRM data and proactively coaches them to sell more, with per-user profile and memory.
• Admin assistant: suggests improvements, changes, and automations to the administrator.

🔁 Two-plane architecture

A control plane (panel + decision queue on Neon) and a reasoning plane (Claude + MCP with scheduled runs that pull the queue, work, and push results back). Every action is logged, with token usage metered per agent.

🛡️ Autonomy with guardrails

Design principle: propose-only. The agent can analyze, draft, and prepare — but no change touches production without a human-approved proposal. That combination of autonomy + control is what makes it safe to trust agents with daily operations.