Guides
Explanations of the parts teams get asked to make decisions about — tokens, agent memory, model size, integration platforms. Written to be read once and referred back to.
API Integration Platforms, Compared: Three Products Sold Under One Word
Zapier, Merge and Temporal get compared as if they were the same purchase. They are three different products. How to tell which one you are buying, and how to model the bill before you sign.
What an AI Token Actually Is, and Where They Accumulate
A token is not a word, and it is not a property of your text. It belongs to the model, which is why the same sentence gets more expensive when you upgrade. Here is what a token is, every place in a request where tokens accumulate, and what the meter costs at the level of a data centre.
A Practical Guide to AI Agent Memory: Building It, Using It, Keeping It Honest
Agent memory is not a brain. It is text you re-send, selected by code you wrote. What the four memory types actually are, where to store each, what replay costs, and the maintenance work nobody budgets for.
Agent Logs Are Not Audit Logs: Monitoring AI That Can Reach Your Secrets
When an agent holds credentials, the attack looks like normal traffic. What standard observability misses, why the log store becomes the liability, and the minimal monitoring set to put in place.
Where Your AI Bill Actually Goes (and How to Cut It Without Switching Vendors)
Most of the overspend on LLM APIs is self-inflicted: tokens you re-send, retries you never cap, a frontier model doing clerk work, and caching discounts you left off. None of it shows up until the invoice, because almost nobody measures token spend per feature. Here is where the money goes, and the order to fix it.
Model Size Is a Deployment Axis
Large and small language models aren't competitors ranked on one axis — they're different tools. Match the class to the task's shape and 'which is better' becomes a routing table. A per-use-case verdict, with the benchmark numbers, including the ones that expired.
Loops, Precisely: What to Automate, What to Queue, and What It Costs You
Agent loops pay only where the work recurs and a machine can verify it — otherwise you have built a queue, not a loop. The taxonomy of the five primitives, a flaky-test loop start to finish, and the review-capacity data behind the argument.
How to Prevent AI Data Leaks: A Practical Guide for Engineering Teams
AI assistants can leak your data through prompt injection. Six rules to prevent AI data leaks, what each one stops, and a checklist for your team.
WEEKLY NOTE
One note per week.
One short note from current work plus 2–3 outside links worth your time.