The Dual-Agent Builder's Starter Workbook — Design AI Agents That Check Each Other's Work

A single AI agent doing real work grades its own homework. It plans, executes, and evaluates its own output — and ships confident mistakes, with no pause point and no independent record of what happened.

The Dual-Agent Builder's Starter Workbook teaches the fix as a design method you fill in for your own project: two specialized agents with different jobs, a supervisor layer that routes work and can pause or stop everything, approval gates between deciding and doing, mandatory independent cross-verification, a four-layer memory model, and an audit log that can prove what the system did and why.

This is a workbook, not an essay. Across 9 chapters you'll complete 9 worksheets — your failure modes, task lifecycle, routing rules, approval-gate matrix, verification rubric, memory design, audit log spec, failure-mode table, and a one-page capstone architecture sheet — plus a 5-test smoke protocol (including the planted-error test that tells you whether your verification is real or theater) and a launch checklist.

Important, stated plainly: this workbook contains no software and no source code. It sells the method — the worksheets, protocols, and design templates — which you then implement in whatever stack you already use. It makes no claims about what a system built with it will earn or achieve.

If you're building agents that touch real files, real messages, or real money, this is the supervision design homework, done in the right order.