Why this notebook exists

Most of what I learn about agentic coding does not fit the paper format I use in my research work. A benchmark result is superseded within a month, a command-line flag changes between releases, and a technique that saved me an afternoon is too small to publish and too useful to lose. The cost of finding these things out is low: at the $2 per million input and $10 per million output token rate this site’s comparisons assume, an afternoon spent driving an agent through a real task works out to a few dollars of tokens, cheap enough that what is worth recording is the technique rather than the bill. The paper format, however, taught me a working habit that transfers well: date every claim, source every number, and write down what failed along with what worked. This notebook applies that habit to my daily work with coding agents, open-weight model deployment, and the cost side of running both.

The rules the site follows are threefold: (1) every entry carries its date, and the longer guides receive a changelog line when they are updated in place; (2) every number links to the source that was fetched before the number was written down, and collected benchmark results are labelled as collected, never presented as my own runs; (3) there are no sponsors, no affiliate links, and no analytics, so no part of these pages is optimized for anything except being checkable.

It should be noted that the entries are drafted with the same coding agents this notebook studies. The pipeline is plain: I give the agent a direction and my research notes, it drafts the entry against the rules above, every number is checked against a source fetched while writing, and I read every line before it publishes. That arrangement is itself part of the experiment, and when it produces something worth examining, it becomes an entry.

The notebook uses two formats. Notes are short and dated, 200 to 800 words, written when something breaks or surprises me. Guides are longer references, 2,000 to 5,000 words, updated in place as the field moves; setting up Claude Code and agentic coding best practices are the two anchors. TODO, dated July 2026: revisit the cost figures across the guides when Claude Sonnet 5 introductory pricing ends on August 31, 2026, since several comparisons assume the $2/$10 per million token rate.