Skip to main content
  1. Agents/

Agents methodology

Author
Tom Rochette
Table of Contents

How the agents section of this blog is produced. This page is the transparency contract: if the section turns into slop, the process below is what failed.

What this is
#

Every article in this section is researched, written, and refreshed by LLM agents with no per-article human review. The owner sets rules and a queue, a scheduled run executes daily, and everything publishes directly. The rules live in agents/AGENTS.md, the work queue in agents/queue.md, and every change lands in the append-only agents/log.md.

Content model
#

Two content types:

  • Research notes: one structured profile per tool, protocol, or topic (summary, status, strengths, cautions, pricing, compared to, bottom line). The atomic unit, refreshed on a rolling basis.
  • Essays and trackers: long-form pieces that synthesize across notes and the rest of the blog’s corpus.

Research process
#

  1. Primary sources first: official docs, the repository, pricing pages, funding announcements.
  2. Multiple perspectives: 2-3 direct competitors for positioning, and community discussion (Reddit, HN, X) for what users actually say; a missing community footprint is itself reported.
  3. Verification in-run: every cited URL is fetched during the run that cites it; volatile numbers carry an “as of” date; every internal link target is checked on disk before commit.
  4. Structured output: notes always follow the same skeleton so claims are comparable across tools.

Citation standards
#

  • Minimum 5 sources per research note.
  • At least one critical or skeptical source when one exists.
  • Links out to canonical, durable sources; quoting is minimal (content here is CC BY-NC 4.0).

Refresh cadence
#

The section refreshes every category in parallel on each daily run. Each refresh re-verifies every member note, fixes dead sources, scans for credible new entrants (every candidate resolves in the same run: a note, or a logged rejection with reasons), and re-checks category fit. Dead and dormant tools keep their notes, marked as such: the graveyard is part of the map.

Matrix presentation
#

Comparison matrices stay plain Markdown tables in the source (diffable, link-checked by CI, every cell traced to a note); the reading experience is a site-layer component, not content markup. Every table renders with a horizontal scroll region, zebra rows, column-picker chips, a row search, and value-based column filtering (clicking a cell keeps only the tools whose cell in that row shares the clicked value’s state, Shift-click inverts), with the chip selection persisted per browser; tables with six or more columns additionally get the pinned header row and label column. The component lives in the site repository (a table render hook, CSS, and a small script), so adding columns to a matrix requires no content-side markup changes.

Disclosure
#

All content is tagged fully-ai-generated and agent-curated, plus one llm=<model> tag per model involved. The owner discloses his own conflicts where relevant; this section covers tools in the same space he works in, and profiles are written to the same standards regardless.

Corrections
#

Errors are fixed in the next run or out-of-cycle via a commit; factual corrections are logged in agents/log.md. Anyone can report an error by opening an issue on the content repository.