CASE STUDY
Broker Outreach Agent
An approval-gated outreach agent. It drafts personalised broker emails with Claude, queues every first email for a human to approve, schedules follow-ups, classifies replies, and writes every action to an audit log.
- My role
- Sole designer and builder of the agent.
- System status
- Built and verified end to end against real services. Runs in a safety mode with dry-run on and follow-up auto-send off.
- Scope
- Approval-gated broker outreach: draft, human-approve, send, schedule follow-ups, classify replies, and log every action.
- Key constraint
- No first-contact email sends without a human approving it.
- Verifiably working
- CSV import, Claude drafting, approval queue, AgentMail send, 7/14/30 day follow-ups, reply classification, weekly report, and audit log. Ten verification steps run against real Claude, Supabase and AgentMail; two bugs caught and fixed.
- Mocked or unmeasured
- No outreach volumes, reply rates or deliverability figures are claimed. Auto-send for follow-ups stays off until the operator turns it on.
Overview
Broker outreach is repetitive, and the risk is sending the wrong thing at scale. This agent keeps a human in the loop on every first contact. It imports broker records from CSV, drafts a first-touch email with Claude, and holds it in an approval queue. Nothing goes out until a person approves it.
Approved messages send from a dedicated AgentMail inbox. The agent then schedules 7, 14, and 30 day follow-ups, polls for replies, classifies them, and escalates anything positive or unclear. It drafts a suggested reply for the human rather than sending one itself, and produces a weekly pipeline report.
What it does
- Imports broker records from CSV
- Drafts personalised first-touch emails with Claude
- Queues every first email for human approval; nothing sends without it
- Sends approved messages from a dedicated AgentMail inbox
- Schedules 7, 14, and 30 day follow-ups
- Polls replies and classifies them, escalating positive or unclear ones
- Drafts suggested replies without auto-sending them
- Produces a weekly pipeline report
- Writes every important event to an audit log
Stack
Verified end to end
This was checked against the real services, not mocks: real Claude, real Supabase, and real AgentMail. Ten verification steps were run, from CSV import through to the weekly report. Two real bugs were caught and fixed in that pass: a silent upsert failure and an over-strict send guard.
Current safety mode keeps dry-run on and auto-send for follow-ups off, so every outbound action stays under human control until the operator chooses otherwise.