Band of Agents: an Honest Lead-Qualification Desk

Lead-ops teams drown in inbound, and two failure modes dominate: the manual cost of gluing sourcing, scoring and deciding together by hand, and silent over-claiming — automated scorers reporting "qualified, high confidence" on evidence they never actually had. In a revenue pipeline, an unaccountable "yes" is worse than a slow "maybe". Band of Agents is a Band room where five specialized agents qualify inbound leads against an Ideal Customer Profile. A Planner is the sole state-machine driver and routes every hop by @mention; a Sourcer pulls candidates; a Judge (an LLM agent on Anthropic Claude) emits a structured verdict with evidence; a Decision agent applies, rejects or parks. All coordination happens in-room as Pydantic-validated stage events — no out-of-band glue. The differentiator is a separate, deterministic governance agent — the Auditor — that audits the others' HONESTY rather than their taste: every evidence item must resolve to a real candidate field, and a high score must actually cite evidence. A red-team fixture forces a score=95 / no-evidence verdict through the same pipe, and the Auditor BLOCKS it deterministically, never hand-inserted. This mirrors Band's own thesis of built-in governance. The pipeline runs reproducibly end-to-end and is covered by 50 offline tests; the Band REST transport and a live Claude Judge are verified against a real room.