Netresearch Agent Engineering

Static, no-build landing page for the Netresearch Agent Engineering System — how real agent sessions become durable, verifiable, distributable engineering knowledge (retro → skills → checkpoints → harness → distribution).

Live: https://netresearch.github.io/agent-engineering/ (German) · https://netresearch.github.io/agent-engineering/en/ (English)

Files

• public/index.html — complete German page, styles, search, theme switching, JSON-LD, local Chrome LanguageModel Q&A and deterministic search fallback
• public/en/index.html — English version, same structure; keep content in sync with the German page
• public/og-image.png, public/og-image-en.png — social previews, rendered from scripts/og-template*.html (npm run og)
• public/robots.txt, public/sitemap.xml
• public/llms.txt — compact machine-readable knowledge summary

No external JavaScript, CSS, analytics, tracking or font request. The CSS declares Netresearch's Raleway/Open Sans typography with system fallbacks.

Local AI behavior

The Q&A feature uses globalThis.LanguageModel when Chrome exposes the built-in Prompt API. It checks availability() without creating a model session; a session is created only after a visitor submits a question. The system prompt contains a compact extraction of this page and requires answers to use that source only. If the API or model is unavailable, the UI does not fake an AI response — it returns the highest-scoring relevant sections of the same page instead.

Verification

python3 scripts/verify_site.py          # markup-decidable gate
npm ci --ignore-scripts && npm run axe  # WCAG 2.1 AA, real browser, light + dark

verify_site.py checks: one h1, heading order, unique IDs, accessible controls, canonical/OG/JSON-LD, no broken internal links, no third-party requests, and that robots.txt, sitemap.xml and llms.txt agree with the page about the base URL and section anchors.

Deploy

Every push to main runs both gates and publishes public/ via .github/workflows/pages.yml (org-reusable pages-build/pages-deploy workflows). A red gate blocks the deploy.