Claude Agents at Work
How to Scope, Connect, and Trust a Fleet of AI Agents
by Isaac Otu
You already use Claude for individual tasks. This book is about the harder problem that starts when you try to run standing parts of a real operation through several agents at once, and still trust what they report back.
Fleets of specialist agents fail in a small number of recurring, recognizable ways: a role goes stale and nobody notices, a handoff drops in silence, a validator passes clean while the output is wrong, an agent states a guess in the same confident voice as a checked fact, a number about your own results that means less than it looks. Each one is named plainly, shown from a real incident in a working fleet that coordinates several ventures with no hired team, and paired with the check that actually catches it.
The argument underneath is deflationary, and the author thinks it is correct: more agents is not more capacity. The skill is scoping each role so you can trust it, connecting them so nothing falls through the gaps, and checking the work so mistakes stay small. Past a certain size, what caps a fleet is not the agent count but one person's ability to keep verifying what it produces.
You will come away able to:
- Scope a role so its boundary is something the agent cannot cross, not an instruction it can talk past - Spot the handoff points where coordination silently breaks, and fix the structure instead of adding a reminder - Read a clean automated pass, a confident claim, and a flattering metric as the narrow things they are - Decide what belongs on a standing schedule and how to know it still runs - Run several separate operations through one fleet without their rules bleeding together - Size a fleet to what you can actually check
It includes a fill-in role-design template, a build-order checklist, and a one-entry-per-failure quick reference.
This edition tightens the argument throughout, tells the incidents in more detail, and adds the template, the checklist, and the failure-mode reference above.