Stewards That Build Stewards

What this book is

A workforce that expands from what actually happens

A multi-agent system starts small on purpose. One agent watches for errors, one drafts commit messages, one answers a narrow question about deployment. That works until the number of distinct, recurring tasks outgrows the number of agents you're willing to hand-build — past that point, growth is bottlenecked on your attention, not on demand. A pattern watcher detects recurring work. A gap detector decides which patterns deserve a dedicated specialist. A stub forge generates a safe, working code skeleton for it. A sign-off gate reviews everything before anything runs unsupervised. None of the four is impressive alone — chained in that order, they buy a system whose capability grows in proportion to actual demand.

This is not a demo built to look clean. Every number, every worked example, and every failure mode comes from a real deployment, including the parts that don't look clean: a discarded false-positive pattern, two independent thresholds that were never reconciled into one number, and a real batch of approved code sitting undeployed because the pipeline generates and reviews faster than a human signs off on the last step. Understanding why that queue exists — and why the human step causing it is a feature, not a missing piece of automation — is the actual lesson underneath the whole pattern.

What you’ll learn

Nine chapters, from pattern detection to the trust ceiling

• the-pattern-watcher — What a recurring task actually looks like before it becomes a candidate for a new agent.
• the-gap-detector — Turning a candidate pattern into a proposal for a new specialist, and discarding the ones that are just noise.
• the-threshold-problem — Two independent thresholds, guarding different failure modes at different costs, that were never reconciled into one number.
• the-stub-forge — Generating a safe, working code skeleton for a proposed specialist's duties.
• the-sign-off-gate — Reviewing generated code before anything runs unsupervised, and why the resulting undeployed queue is a feature.
• message-passing — The coordination problems inherent to asynchronous handoffs between independent agents: crash recovery, duplicate delivery, message ordering, backpressure.
• what-goes-wrong — A real discarded false-positive pattern and what the discard reveals about the pipeline's judgment calls.
• the-trust-ceiling — What determines when — and whether — a human should be removed from a deployment loop at all.
• conclusion — The minimum viable version of the whole pattern, and what it buys you at any scale.

From the book

A preview

A real batch of reviewed, approved code sits un-deployed, because the pipeline generates and reviews faster than a human signs off on the last step. That is not a flaw hidden in a footnote. It is the subject of two chapters.

Who it’s for

Operators already running a multi-agent setup

This book assumes you already run a multi-agent Claude Code setup with more than one or two specialist agents, and that you've felt the specific friction of deciding whether a recurring task justifies a new agent or just another ad hoc prompt. It does not teach multi-agent basics or Claude Code fundamentals — it's the honest accounting of one specific architectural pattern: a workforce that grows itself.

A longer excerpt is available to newsletter subscribers.

Buy Ebook on Amazon