MultiAgentEDUstack

A working implementation of the multi-agent curriculum pipeline described in Curriculum at Model Speed, a system that sources AI knowledge fast enough to matter, curates it against a five-tier credibility rubric, and scaffolds it into curriculum, without a six-month lag.

Curriculum at Model Speed: design essay and source table. The screenshot is the live desk that implements it.

This is a real, running system, not a design doc. Everything below is either verified working against live data, or explicitly marked as a stub with the reason it isn't.

Published output

Pipeline markdown is synced to GitHub under published/, by category and date:

Category	Path
Digests	published/digests/
Forecasts	published/forecasts/
Wiki	published/wiki/
Curriculum	published/curriculum/
Labs	published/labs/
Decay scans	published/decay/

Layout notes: published/README.md.

Architecture

Five scouts (scheduled nightly, no LLM call)
  arxiv_scout · hn_scout · github_trending_scout · blog_scout · youtube_scout
  (+ reddit_scout, social_scout: real adapters, credential-gated, see below)
        │
        ▼
raw_items (SQLite)
        │
        │  pipeline/dedupe.py -- rule-based tier scoring + story merging
        ▼
curated_items
        │
        │  .claude/skills/synthesis-digest -- daily, via headless Claude Code
        ▼
published/digests/YYYY-MM-DD.md  +  topics on curated_items
        │
        │  .claude/skills/weekly-wiki -- Sundays, rolls daily digests
        ▼
published/wiki/YYYY-MM-DD.md
        │
        │  .claude/skills/trend-forecast -- Sundays, reads assigned topics
        ▼
forecast_watchlist  +  published/forecasts/YYYY-MM-DD.md
        │
        │  .claude/skills/curriculum-scaffold -- run by hand, per topic
        ▼
curriculum_units  +  published/curriculum/YYYY-MM-DD/*.md
        │
        │  .claude/skills/lab-generation -- run by hand, per unit
        ▼
lab_specs  +  published/labs/YYYY-MM-DD/*.md  (spec only, no live cloud)
        │
        │  .claude/skills/editorial-review -- human-gated, never auto-approves
        ▼
editorial_reviews  →  shipped curriculum

.claude/skills/decay-deprecation runs independently, flagging shipped units that have gone stale.

What's actually running right now

Three systemd user timers (plus a deprecated synthesis alias):

• multiagentedustack-ingest.timer: 02:15 daily. Runs all seven scouts and the dedup pass. Deterministic, no LLM call, no cost. Offset from the author's existing open-brain-compile.timer (02:30) to avoid contention.
• multiagentedustack-digest.timer: 03:00 daily. Runs synthesis-digest headlessly via claude -p for new-article summaries.
• multiagentedustack-weekly.timer: Sunday 03:30. Runs weekly-wiki then trend-forecast headlessly via claude -p.

If you still have the old Sunday-only unit enabled, migrate:

systemctl --user disable --now multiagentedustack-synthesis.timer
systemctl --user enable --now multiagentedustack-digest.timer
systemctl --user enable --now multiagentedustack-weekly.timer

Check them: systemctl --user list-timers 'multiagentedustack-*' Logs: logs/ingest-*.log, logs/digest-*.log, logs/weekly-*.log.

curriculum-scaffold, lab-generation, and editorial-review are not on a timer, they're judgment calls worth being in the loop for. Run them by hand:

claude -p "/curriculum-scaffold" --allowedTools "Bash Read Write Edit"

The database

SQL is the source of truth for structured state; markdown under published/ is a regenerable teaching view (same "wiki is a build artifact" discipline), organized by category and date and synced to GitHub by scripts/publish-output.sh after daily/weekly (and manual) generation. Schema: db/schema.sql. The live database (db/maes.sqlite3) and local logs stay gitignored.

Everything an agent skill reads or writes goes through scripts/db.py (python3 scripts/db.py --help for the full command list) rather than freehand SQL scattered across prompts.

Source database

data/sources.yaml is the credibility-tiered source list (55 sources, 9 categories) from the companion blog post's interactive source explorer. Scouts read their target lists from it (YouTube channels, blog feeds); arXiv categories and HN/GitHub keyword filters are set directly in their scripts.

What's real vs. stubbed, and why

Piece	Status	Why
arXiv, HN, GitHub trending, blog/newsletter RSS scouts	Real, verified live	Public APIs/feeds, no credentials needed
YouTube scout	Real, verified live	yt-dlp for channel listing + auto-caption extraction, no YouTube Data API key needed
Reddit scout	Real adapter, credential-gated	Reddit's unauthenticated .json endpoints 403 from this network (verified directly, not assumed) since their 2023 API lock-down. Needs a free OAuth app (REDDIT_CLIENT_ID/SECRET in .env)
Bluesky scout	Real adapter, unverified from this network	Free (handle + app password, no paid tier), but this sandbox's network got blocked by Bluesky's edge while building this. Should work from a normal residential/office network
X/Twitter scout	Real adapter, inert without a paid API tier	X's v2 API requires paid access for meaningful read scope as of 2026; this repo doesn't assume you're paying for it
Credibility + Dedup	Real, rule-based	Tier lookup by scout/category, exact-URL then title-similarity merge. No LLM call
Synthesis-Digest (daily), Weekly-Wiki + Trend-Forecast (Sunday)	Real, LLM-driven, scheduled	Verified end to end via headless claude -p against live data
Curriculum-Scaffold, Lab-Generation	Real, LLM-driven, manual	Judgment calls, run by hand rather than unattended
Editorial-Review	Real, human-gated	Produces the two-axis report and explicitly refuses to auto-approve anything; a decision is only recorded after a human states one
Lab provisioning (live cloud sandboxes)	Not built	lab-generation produces a spec only. This repo has no cloud credentials wired in, and standing up billed infrastructure isn't something an unattended agent should do on its own judgment. Turning a spec into a real environment is a separate, explicitly-authorized step you take by hand
Behavior-change telemetry	Scoped down	The blog post's design targets a ~10,000-engineer org with toolchain-wide instrumentation. At solo-practitioner scale, telemetry_events tracks whether a surfaced item actually gets used (opened, cited in a post, promoted to curriculum), a real but much smaller version of the same idea

Setup

pip install -r requirements.txt
cp .env.example .env   # fill in whatever credentials you actually have
python3 scripts/ingest.sh   # or: bash scripts/ingest.sh
python3 scripts/db.py new-items

See .env.example for exactly which credentials unlock which scouts.