Building AI Agents That Work in Production

Index — All ArticlesNewest First

No. 40Oct 2026
When Your Manager Says to Feed the Data to AI
"Feed the data to AI" is a goal with no spec. Five questions turn it into a decision about users, sources, permissions, freshness, and how answers get checked.
No. 39Oct 2026
An Executive's Guide to AI Pricing
AI services bill by the token, the seat, and the run. How each meter works, why output costs five times input, and how to budget per accepted result.
No. 38Sep 2026
Agent Memory Needs an Expiration Rule
AI agent memory becomes a liability when facts outlast their source. Give each memory a timestamp, invalidation rule, and path back to verification.
No. 37Sep 2026
AI Agent Observability Stops at the Trace
AI agent observability shows how a run happened. Evaluation and outcome checks establish whether its decision produced a useful, correct result.
No. 36Sep 2026
Model Routing Is an Evaluation Problem
Model routing needs acceptance thresholds, real evaluation, and escalation paths. Costs fall only when every route earns an accepted result.
No. 35Sep 2026
Claude 201: A Repeated Task Needs a System
Recurring Claude tasks need durable context, clear completion criteria, and correction loops that improve every run without relying on chat history.
No. 34Aug 2026
The Model Has to Read the Manual
Complex data APIs work better when query tools require a current guide receipt, then return errors that show the model exactly how to recover.
No. 33Aug 2026
Generated Code Is a Build Artifact
Generated code needs assurance through tests, contracts, monitoring, and focused inspection instead of universal line-by-line review.
No. 32Aug 2026
You Want Your Apps in AI, Not AI in Apps
AI products should meet users where work begins, while durable state and detailed controls stay in the application.
No. 31Aug 2026
Claude 101: Three Useful Workflows
Claude 101 starts with a real job, source material, and a result you can judge. Three first workflows for work already waiting on your desk this week.
No. 30Aug 2026
AI Inference Cost Is Architecture
AI systems get expensive when the model owns every branch. Deterministic orchestration keeps reasoning bounded, testable, and worth its token bill.
No. 29Aug 2026
Source Code Has No Value
Source code is necessary, but it is not a durable moat. Customer problem understanding is the advantage a competitor cannot copy from a repository.
No. 28Aug 2026
Rules Are Not Controls
A rules file moves the odds. It does not bind anything. HANDBOOK.md measures the gap between policy and compliance.
No. 27Jul 2026
Carry Your Own Clock
The control group quit and the benchmarks broke. There is no shared instrument for what AI does to engineering work, so carry your own.
No. 26Jul 2026
Responsible Charge
Every field that solved delegated work solved it with a name on a line and a license that can be taken. AI governance is a hunt for a way to skip that step.
No. 25Jul 2026
It Worked Once
A demo proves a thing is possible. It says nothing about how often it works, and production runs entirely on how often. Possibility is not reliability.
No. 24Jul 2026
Children of the Magenta Line
The cockpit ran this experiment first: the work automation spares you is where your judgment was grown. Knowledge work is about to get the same bill.
SeriesJun 2026
Six-Part Series
The User-Code Stack
Your users' assistants already write code against your product. Six parts on running it safely, end to end: the case for allowing it, the execution boundary, the sandbox and policy at the heart, an identity that survives the trip in, a credential broker underneath, and the primitives the whole platform composes from.
No. 17Jun 2026
A Channel Is Not an Assignment
Work assignment is a solved problem, and chat is the tool we know fails at it. Claude Tag hands that job to an agent, which only makes the failure worse.
No. 16May 2026
The Knowledge Rancher
When generation becomes free, the job stops being production and turns into managing a herd you didn't have to raise. What's left is judgment and timing, and that drains you in a way typing never did.
No. 15May 2026
MCP Apps: When Your Server Needs a Face
MCP Apps let servers render interactive UI directly in conversations. Use one when people need to manipulate state that text handles badly.
No. 14May 2026
Writing a Good MCP Server
Most MCP servers expose too many tools with bad names and useless errors. Practical lessons from building servers that models actually use well.
No. 13May 2026
Intro to Claude Context
Claude context has six layers: instructions, projects, skills, memory, chat search, and the current task. Put durable rules in layers you control.
No. 12May 2026
Prompted but Not Read
We used to write 'Dictated but not read' at the bottom of letters when a secretary did the typing. The parallel to what we're all doing now is too perfect to ignore.
No. 10May 2026
Toxic AI Mentality
A frame people fall into where AI failure becomes a trophy. The screenshot is queued before the prompt is sent. Meanwhile, everyone around them is shipping.
No. 11May 2026
We Automate What We Can Verify
Code was the first big win for AI agents because code is cheap to check. The same rule decides which business processes compress next, and which stay stuck.
No. 09Apr 2026
What 'Giving Data to AI' Actually Means
The mental model is wrong. You don't pour your database into a prompt. You equip the model to go look, and that's the project every business is actually working on.
No. 07Apr 2026
Distraction Degrades the Model
Packing competing objectives into one prompt costs quality on both. The intuition for when to split is a real skill.
No. 08Apr 2026
Using AI Is Management
People who've managed teams tend to get more out of AI faster. The skills transfer directly: delegate clearly, define success, don't blame the team member for your assignment.
No. 05Apr 2026
Micromanaging the Model Backfires
Modern LLMs are capable enough that step-by-step instructions constrain them. A destination outperforms a procedure. Declarative beats imperative.
No. 06Apr 2026
The English Trap
You're not having a conversation with an LLM. You're writing a spec for a system that will exploit every ambiguity you leave on the table. The sooner you treat it that way, the sooner your results stop being wildly inconsistent.
No. 04Apr 2026
Agents Are About Correction
AI with tools can act. What makes an agent different is what happens when the action fails.
No. 03Apr 2026
Production Output Belongs in a Tool Call
When you pipe LLM chat responses into production systems, you're using a scratch channel as a delivery mechanism. Tool calls separate reasoning from output and fix the reliability problems you didn't know you had.
No. 02Apr 2026
MCP Is the OAuth Layer
The anti-MCP take is a single-user take. When your agent serves multiple users accessing their data on remote services, you need a protocol. MCP is the one.
No. 01Mar 2026
Why Emphasis Works in Prompts
Bold text and CAPS in prompts aren't voodoo. They work because of how attention literally functions in transformers. Understanding the mechanism makes you a better prompt engineer.

Lessons from production.Things that work, things that don't.

When Your Manager Says to Feed the Data to AI

An Executive's Guide to AI Pricing

Agent Memory Needs an Expiration Rule

AI Agent Observability Stops at the Trace

Model Routing Is an Evaluation Problem

Claude 201: A Repeated Task Needs a System

The Model Has to Read the Manual

Generated Code Is a Build Artifact

You Want Your Apps in AI, Not AI in Apps

Claude 101: Three Useful Workflows

AI Inference Cost Is Architecture

Source Code Has No Value

Rules Are Not Controls

Carry Your Own Clock

Responsible Charge

It Worked Once

Children of the Magenta Line

The User-Code Stack

A Channel Is Not an Assignment

The Knowledge Rancher

MCP Apps: When Your Server Needs a Face

Writing a Good MCP Server

Intro to Claude Context

Prompted but Not Read

Toxic AI Mentality

We Automate What We Can Verify

What 'Giving Data to AI' Actually Means

Distraction Degrades the Model

Using AI Is Management

Micromanaging the Model Backfires

The English Trap

Agents Are About Correction

Production Output Belongs in a Tool Call

MCP Is the OAuth Layer

Why Emphasis Works in Prompts