Book · serialized in public
The Production Agent
How to ship agentic AI to production without getting paged at 3am.
Battle-tested playbooks from shipping multi-agent systems at 99.9% uptime — orchestration, evals, security, MCP integrations, and the failure modes you don't read about in the launch posts.
Chapters 1-2 free on Substack and on this site. Chapters 3-5 are paid Substack subscriber-only (or buy the lifetime PDF bundle on Gumroad).
Table of Contents
5 chapters · 2 free
- Chapter 122 min readFREE
Designing the Production Agent
Orchestration patterns, agent contracts, and the deny-by-default starting line.
How to design an agent that survives contact with production. Tool inventories, capability contracts, deny-by-default surfaces, and the cost router pattern that takes simple problems off the frontier model.
Read draft - Chapter 218 min readFREE
MCP Security & STDIO Posture
Why MCP STDIO is RCE-by-design and the sandbox patterns that fix it.
The Model Context Protocol unlocks composable agent tooling — and ships a critical-by-default vulnerability in its STDIO transport. CVE-2026-30623, what it actually means, and the three-layer sandbox pattern that keeps it from owning your host.
Read draft - Chapter 331 min readPAIDOutlined
Agent Supply-Chain Audit
50-point audit for the upstream surface of an agent stack.
The upstream supply chain of an AI agent is model provenance + MCP server graph + eval-set lineage + secret rotation policy + audit-emit channels. 50 audit points, each with a concrete check and a remediation.
Releases on Substack first - Chapter 428 min readPAIDOutlined
Migrating to Claude Managed Agents
Step-by-step runbook for moving production workloads.
Claude Managed Agents shipped with Anthropic's nine-connector creative launch and ten finance-agent templates. Here's the migration runbook — control-plane wiring, capability-lease bindings, MCP connector hardening, and the rollback playbook for when a vendor surface shifts.
Releases on Substack first - Chapter 524 min readPAIDOutlined
Agent-on-Agent Commerce — The Honest Version
What works today, what's marketing, and what the next 12 months look like.
Anthropic Project Deal, Cloudflare Mesh, Cisco Agentic Workforce Identity, Okta for AI Agents — the agent-on-agent stack is real. Here's what actually ships in 2026 vs what's vendor positioning, and what an honest /api/quote agent contract looks like.
Releases on Substack first
Why this book
Most LLM books teach you to call the API. Most agent posts celebrate a launch. Neither tells you what breaks at week 2 of production, what your audit-emit channel actually needs to capture, or why your MCP STDIO config just gave an attacker a shell on your box.
This is the missing manual — orchestration patterns, eval harnesses, capability-lease identity, MCP-security posture, supply-chain audit points, and the migration runbook for moving onto Claude Managed Agents — distilled from shipping multi-agent systems at 99.9 % uptime, plus the open-source tooling (agent-audit-kit, agent-airlock, mnemo) that I use to keep them honest.
Each chapter ships on The Production Agent first — subscribers get it free if it's a free chapter, paid subscribers get every chapter. Or buy the lifetime PDF bundle on Gumroad once if you'd rather not subscribe.