A small monorepo where I build practical AI agents that actually run somewhere. Research pipelines live in Python, delivery lives on Cloudflare Workers + Telegram.
Right now that's two projects: NewUpdater (tech news digest, 3x daily) and RepoRadar (GitHub release radar, in progress). The layout is set up so adding a third one doesn't mean rewiring the first two.
agent-factory/
projects/
newupdater/ # Python research pipeline (LangChain + LangGraph)
reporadar/ # placeholder, GitHub monitoring coming next
infrastructure/
cloudflare/
newupdater-bot/ # Worker: Telegram webhook, cron digests, Groq summaries
shared/
agent_core/ # placeholder for code every agent reuses
src/ # root package, mostly workspace glue for now
| Project | What it does | Status |
|---|---|---|
| NewUpdater | Collects tech news (Google News, Bluesky, X), ranks it, sends a Telegram digest morning / afternoon / night | Working locally, Worker scaffolded |
| RepoRadar | Watches GitHub repos for releases and meaningful changes, sends a daily digest | Not started |
- Python does the research.
projects/newupdaterfetches sources, dedups, ranks, and can call Groq through LangChain. Good for experimenting. - The Worker does the delivery.
infrastructure/cloudflare/newupdater-botowns the Telegram webhook (/latest,/start), the 3x-day cron, dedup in KV, and the final message formatting. It re-implements the fetch + rank in TypeScript so it can run on its own schedule without my laptop. - Nothing at the repo root runs in production. Each deployable has its own config and lockfile, so deploying the bot can't break the Python side and vice versa.
Prereqs: Python 3.12 + uv, Node 20+ for the Worker.
# python pipeline
uv sync
uv run python projects/newupdater/inspect_tool.py
# worker (from its own folder)
cd infrastructure/cloudflare/newupdater-bot
cp .dev.vars.example .dev.vars # fill in tokens, never commit .dev.vars
npm install
npm run typecheck
npx wrangler devSecrets live in two places and both stay out of git: a .env next to the Python code, and .dev.vars / wrangler secret for the Worker. If you need an example value, commit a .example file, not the real one.
To receive Telegram messages: create a bot with BotFather, get your chat id from getUpdates, then point the webhook at /telegram/webhook. The preview endpoint (GET /digest/preview?topic=AI agents) is the safe way to test without sending anything.
- Copy the shape:
projects/<name>/for logic,infrastructure/cloudflare/<name>-bot/only if it needs its own Worker. - Give it its own
pyproject.toml(orpackage.json), its ownwrangler.tomlif it deploys, and its own KV/naming prefix. Don't import across projects except throughshared/agent_core. - Add a row to the table above and a 5-line section under its folder. Keep the root README short.
- Add the workspace member (
tool.uv.workspace) or install the Worker separately. Root stays glue, never logic.
- One
wrangler.tomlper Worker, nothing deploy-related at the root. No root GitHub workflow that deploys everything at once. - Separate lockfiles:
uv.lockfor Python,package-lock.jsonper Worker. Bumping a dep in one project doesn't touch the others. - KV keys are prefixed per service (
seen:inSEEN_KVbelongs to newupdater-bot; a new bot gets its own namespace or prefix). - Cron schedules live with the Worker that owns them (
0 7,12,19 * * *for the digests). Changing digest times never touches RepoRadar. - The two things most likely to bite later: committing
.dev.vars/.envby accident (both are gitignored — keep it that way), and sharing one KV namespace across bots without prefixes (don't).
- NewUpdater research pipeline runs (Google News + Bluesky + X verified)
- Worker scaffold typechecks (webhook, cron, dedup, Groq summary)
- Create KV namespace, set secrets, first deploy + webhook
- Tune topics and digest times
- Start RepoRadar using the same pattern