Build log / Phase 3
Finishing the brief — pages, sources, and a hundred listings
Compare, best-of, case studies, and the public build log landed. Guides went from two to eight. The directory grew from five seed tools to a hundred-plus researched YAML files. CI split so vendor URLs cannot fail the build.
This is the second public entry. It is still not a launch narrative.
What was specified
Finish the missing page types. Add guides and sourced case studies. Grow the directory toward ~100 listings. Make CI useful. Wire issue-to-PR without auto-merge. Do not invent metrics.
What we actually did
Pages. /compare/[a]-vs-[b] is a curated list, not a cross-product. /best/[query] is a filter on structured fields. /case-studies and /build-log are first-class. The homepage lists every tool and links the new surfaces.
Guides. Four requested topics (intent, verification, sandboxing, measurement) plus two more (self-hosted assembly, choosing a harness). Eight total.
Case studies. Stripe Sessions 2026 (first-party keynote), Shopify River (shopify.engineering), StrongDM’s own factory site, and Uber’s stack with uReview numbers verified on uber.com and Minion percentages dropped because they only appear in recap blogs and talk titles.
Directory. One YAML per product. Windsurf is not a separate slug — Cognition’s own blog says it is Devin Desktop. Cody is not listed next to Amp. Continue is listed as the open extension after the Cursor acquisition, with that risk in the take. Cloudflare Agents and Cloudflare Sandbox are separate jobs. Publisher conflict on Trevize (and a smaller “we use this” note on Cloudflare Agents) stays explicit.
CI. npm run validate is schema, slugs, duplicates — no network. npm run check-links is isolated and continue-on-error in Actions. PR comments use /tools/<slug> unless PREVIEW_BASE_URL is set. factory-draft / workflow_dispatch opens a draft PR. Nothing merges.
What broke or stayed broken
GitHub Actions on a private repo may still not run until the owner enables it. softwarefactory.build DNS may still be missing. This environment still cannot wrangler deploy. Those are operator facts, not YAML bugs.
Third-party URLs will rot. That is why link check cannot fail the build.
What agents still get wrong
- Treating conference recaps as uber.com.
- Listing Windsurf and Devin as two products after the rename.
- Printing sandbox millisecond leaderboards from vendor comparison blogs.
- Wanting to auto-merge the draft listing PR.
The listings are editorial and dated 2026-08-19. They are not a census of every GitHub repo with “agent” in the README.