Reading · Agent methodology
Reading list
Work I return to when thinking about how agentic systems are built and measured, organized into four tracks. Each track pairs a small set of foundational papers with an automatically refreshed view of recent work, so the list stays current without pretending that everything recent is important.
Loading reading list…
How this list is built
The selection logic is deliberately simple and kept in version control at scripts/fetch_papers.py, so the ranking can be inspected rather than taken on trust.
- Foundations are hand-picked, metadata is not. I choose the arXiv identifiers; titles, authors, and abstracts are always fetched from the Hugging Face papers API. Nothing is transcribed by hand, so a citation cannot silently drift from the real paper.
- Recent work is filtered, not mirrored. Candidates come from the daily feed plus targeted searches, then pass a keyword model with per-track weights, penalties for off-domain topics, and a minimum score. Most candidates are rejected.
- Relevance ranks above popularity. Upvotes break ties; they never drive the ordering. A widely shared paper that does not match the track does not appear.
- The automated half carries no endorsement. Foundations reflect a considered judgment, and each carries a note on what it establishes and what it leaves unsettled. The recent section is a filtered view of what is being published, deliberately left unannotated, and inclusion is not a claim that a paper is correct or important.
Where the automated and hand-judged parts of this site divide, and how to verify either, is set out in the colophon.
Commentary reflects publicly available material and general methodology only, and represents my own views rather than those of my employer.