Free field guide · Book II

So your agents ship code. Now make that boring.

This field guide is the full story of how we run AI agents as production software: the contract that feeds them, the gate that authorizes them, the failure paths designed for them, and the discipline that lets a security finding dispatch its own fix.

12 chaptersReal production system

Get the field guide
web edition, free

The complete guide: the self-healing loop, the day it broke, the work contract, the Health Code, designed failure, verification, observability, memory, the new attack surface, and the 24-question scorecard.

How it starts

The vulnerability fixed itself.

A security scanner finds a vulnerable dependency. Nobody reads the alert, because nobody has to: the finding becomes a validated work order, passes an authorization gate that recognizes a security finding as its own justification, and a sandboxed runner opens the pull request.

The only human in the loop is the one with the most leverage: the reviewer who merges. This guide is everything that makes that loop safe to run.

front doors, one work contract

authorization gate, fails closed

merges an agent can perform

jobs allowed to vanish silently

One self-healing loop proves possibility. An operation needs the loop to hold when things go wrong.

The problem this guide solves

Capability is what your agents can do. Operations is what they cannot.

Anyone with an API key can rent the capability. The operation is the part you have to build. If your agents write code, run jobs, or touch production today, run this test against your own setup:

You probably have

capability

Agents that write and ship real code
Automations wired to your tools
Impressive one-off saves to talk about
Careful people watching over it all

You probably don’t

operations

One gate deciding what agents may change
A designed place where failed work lands
Checks that catch tests asserting nothing
An audit of what agents run automatically

Inside the guide

Twelve chapters. One argument.

01The fix nobody wroteA security finding dispatches its own remediation.

02The day it was not boringThe broken deploy that wrote our rules.

03Run the kitchenAI Ops defined, and why ours has kitchen names.

04Six doors, one envelopeThe work contract every trigger speaks.

05The Health CodeWho may change what, drawn as one matrix.

06Every failure gets an addressDead letters, replay, exactly-once.

07Never trust the authorQuality gates that ignore who wrote the code.

08Agents appear in the tracesObservability for the agents themselves.

09Memory as infrastructureA governed store, and the two-memories war story.

10The new attack surfaceThe worm that targeted agent config files.

11What this does not solveThe honest limits, load-bearing on purpose.

12The scorecard24 questions to run against your operation.

Who this is for

Written for you if

Your agents already do real work: writing code, running jobs, touching production systems
The operation around them grew by accident, and you know it, and nights before releases feel that way
You could not say, mechanism by mechanism, what stops a bad agent change tonight
You want the discipline without a platform team to build it for you

Skip it if

You’re deciding whether to try AI at all
Your agents only draft text a human always reviews
You want prompt-engineering tips
You need a basic CI/CD tutorial

If the honest answer is a person’s name instead of a mechanism, you have an AI operations problem.

Get the field guide Book an Agentic Workflow Product Review