Book Six/Make AI Take Action
Production AI Agents
Build, Trust, and Ship Agents That Do Real Work
A chatbot tells you the refund policy; an agent issues the refund — and that one word, "issues," changes everything about how you build and trust it. Give a model tools and a loop, connect it to the world with MCP, make it survive crashes, then earn the right to ship it: human approval on dangerous actions, guardrails against prompt injection, evals that prove it works, tracing that shows what it did, and cost control that keeps it viable. Every API verified against a compiling project.
$39139 pages · fixed-layout PDF · every code sample verified
Instant PDF download + read online · lifetime access · paperback coming soon
What’s inside
- Build a tool-using agent in twenty lines, then master the tool craft that decides whether it succeeds
- Connect to the world with MCP — consume any server, and publish your own tools as one
- Steer the loop, split work across subagents, and make long-running agents survive crashes and deploys
- Earn trust: human approval, prompt-injection defense, evals, tracing, and cost per completed task
Tools · The LoopMCP · IntegrationsDurable · Multi-agentApproval · GuardrailsEvals · Tracing · Cost
How it’s organized
- 1What an Agent Istools, the loop, tool craft
- 2Connecting the WorldMCP client, MCP server, patterns
- 3Orchestrationsteering, subagents, durability
- 4Trustapproval, guardrails, memory
- 5Prove It Worksevals, tracing, cost
- 6Ship & CapstoneCI gates, deploy, the Cortex Agent
