Guides
Durable, maintained references on building agent systems that behave in production. Updated as the ground shifts.
A Glossary of Agent Architecture
Forty-two terms defined precisely enough to argue with, including which ones get routinely conflated and why the distinction matters.
Agent Failure Modes: A Taxonomy
Hallucination is not a useful category. Twenty-one ways agent systems actually break in production, grouped by where the failure originates.
Buy, Build, or Wait: Evaluating AI Capability as a CTO
The decision before tool evaluation. Three options, not two — and in a fast-moving market, choose for cheap reversal rather than best expected value.
Hallucination Is a Data Contract Problem
Trace a wrong number back and you rarely find a model inventing it. You find an undefined term, an unstated grain, an ambiguous null, or a stale snapshot.
How to Evaluate an AI Tool for Your Business
The scores in any comparison expire. The method does not. How to run the evaluation yourself and reach a decision you can defend in six months.
Letting an Agent Touch the Warehouse Safely
Read replicas, query budgets, row limits and closed egress paths — the layers that turn an arbitrary query from a non-deterministic caller into a bounded operation.
Making an Agent Explain What It Did to Your Data
A generated explanation is not evidence. Capture the path mechanically — query, as-of, row counts, definitions — and assemble the explanation from the record.
Shadow Mode: How to Deploy an Agent That Can't Hurt You Yet
Real production traffic, real decisions, no consequences. How to run an agent in shadow, how to grade it, and how to set the promotion threshold before you look at the data.
Technical Debt in the Age of Generated Code
Writing code was never the constraint. Generation collapsed the cheap half of the problem and left comprehension exactly where it was.
The Agent Reliability Handbook
What it takes to run an agent in production without it quietly breaking things. Contracts, blast radius, replay, and the checklist before you give it write access.
Threat Modeling an AI Agent
Conventional threat models assume data and instructions stay separate. Agents dissolve that boundary by design. Here is how to model one properly.
Tool Contracts: The Interface Between a Reasoning System and the Real World
A tool description is the entire world model a reasoning system uses to decide whether to fire an irreversible action. The nine things a contract has to declare.
What Horse Racing Taught Me About Model Confidence
Being right is not the goal. Knowing how often you are right is. Calibration, edge, staking, and the discipline of the pass.
What to Ask Before Approving an Agent Deployment
Nine questions whose answers are hard to fake, each with what a real answer sounds like and what a hand-wave sounds like.
Why LLMs Are Bad at Big Data
Language models produce plausible text, not correct numbers. Why bigger context windows do not fix it, and the architecture that does.