evidence-backed guidance for coding agents in production software, from individual projects to startups and big tech.
read the publication or go directly to a guide:
| guide | focus |
|---|---|
| codex | cli, desktop, ide, cloud, worktrees, permissions, and durable configuration |
| claude code | terminal, desktop, ide, skills, hooks, agents, and permissions |
| shared operating system | authority, isolation, verification, handoffs, and review |
| market and hardware | surfaces, harnesses, models, orchestration, and local-resource economics |
the guide separates four states:
hands-on: reproduced by the author with the product and date identified.source-verified: checked against a linked primary source.inference: a reasoned recommendation built from stated evidence.unknown: a material gap that remains visible.
the field lab publishes task specifications, sanitized run records, checks, artifacts, redactions, and limitations. current claude code guidance remains source-verified until the paired hands-on run is complete.
docs/contains the canonical markdown, source registry, decisions, and field-run records.src/renders the astro and starlight publication.plugins/andhooks/are frozen compatibility surfaces through 2026-11-05..github/contains deterministic validation, read-only freshness checks, and the github pages workflow.
bun install --frozen-lockfile
bun run check:field-runs
bun run check
bun run build
bun run test:site
bun run test:a11ylegacy plugin paths and their final-support policy are documented at agents.anipotts.com/legacy.
built and maintained by ani potts. corrections with primary sources and reproducible field evidence are welcome.
MIT