Three agents, one principle
Renewal Agent, Pipeline Agent, and the AI Workbench each call their own API route (/api/agent, /api/pipeline-agent, /api/ai) against the chamber's live data: real tool-calling, not a static script pretending to think. Every one of them stops short of the last step: they propose, a staff member decides.
What each one actually does
Renewal Agent scans the at-risk member queue, pulls context per member (engagement, dues status, event history, referrals), and hands back a prioritized worklist with a drafted email per member. Pipeline Agent does the equivalent for open prospect deals, ranking by value at stake and how recoverable a stalled deal looks. The AI Workbench covers six narrower staff tasks: summarizing meeting notes, extracting a structured record from a pasted email, drafting meeting minutes, building a member profile report, drafting outreach copy, or a free-form question against the workspace's own data.
Agents draft. People send.
Every agent-drafted email, note, or worklist item requires a person to act on it: Renewal Agent drafts wait in a Draft inbox for a person to approve, and Pipeline Agent and Workbench drafts are copied by staff. To send any of them, a person pastes it into their own email. None of the agents can send communications on their own. Actions taken from an agent's suggestion are logged, so there's a record of what was proposed versus what a person actually decided to do with it.