🤖 Agentic AI Engineering Practice flagship
What built cv-deck.pages.dev ·
what wrote the resume you are reading ·
what runs every commit on staging under guardrails I authored.
- What it is — a working practice for letting an AI agent own the production codebase while a human owns the trigger. The two cards below are live demonstrations, not separate projects.
- Dev box — Linux 6.17, Python 3.11 venv (PEP 668 isolated), Node 20, Git, wrangler v4.100, Playwright headless Chromium.
- AI orchestration — Hermes subagents for parallel reasoning; LangGraph-style delegation in the agent layer;
AGENTS.mdre-read on every session start. - State & data — Cloudflare D1 + Drizzle ORM, R2 for assets, KV for session config, JWT auth, role-based access control.
- Hosting & delivery — Cloudflare Workers (API), Pages (SPA), Cron Triggers, GitHub Actions for CI,
wrangler pages deploywith--confirmgating. - 3 Critical Rules in
AGENTS.md— no auto-promote to prod · explain before answering · no DB change without explicit instruction. Violating one fails the agent's own review. - Dry-run +
--confirmgating — every mutating command (deploy, DB write, file move) prints a plan and stops. Audit log at~/.hermes/d1-audit.logon every confirmed write. - 5-source verification chain —
git log+git reflog+git ls-remote+ CF Pages REST API +curl-bundle grep. All five must agree before a release is tagged. - Incident-response — caught a fabrication incident mid-session, acknowledged immediately, built detection so it cannot recur.