The Self-Improving Agent Playbook — Stanford's $4,725 CS329A Course, Mapped, Plus 4 Techniques You Can Run Today

The companion to the video. Stanford charges $4,725 for the paid version of CS329A — Self-Improving AI Agents — and enrollment is currently closed with a waitlist. The lectures just landed free on Stanford Online's YouTube channel: 9 core lectures, ~10.5 hours, taught by Azalia Mirhoseini (Stanford, ex-Anthropic, ex-Google Brain) and Aakanksha Chowdhery (led Google's 540B-parameter PaLM). INSIDE: (1) THE VERDICT BOARD — all 9 lectures stamped WATCH / IF / SAMPLE with the exact condition for each and direct links, including the one lecture where you should skip straight to minute 46:31. (2) THE 3-LECTURE SPINE — ~3.5 hours that carry most of the practitioner value, plus the six stats worth remembering with their precise context (16%→56% at 250 samples; task horizons doubling every 7 months; the 59-min vs 15-min reliability cliff). (3) FOUR STEAL-TODAY TECHNIQUES AS COPY-PASTE PROMPTS — Stop Asking Once (sample-verify-FUSE, where fusion beat the best single answer), Feedback Beats Genius (the execution-feedback loop with the hidden-test trick), Never Trust One Judge (different-model critics + the judge-audit prompt), and The Contractor Rule (the 15-minute delegation checklist). No GPUs, no fine-tuning — everything works in Claude Code, Codex, Cursor, or plain API calls. (4) THE 2-WEEK PLAN — one hour a day from tonight. Sourced from the full lecture transcripts on upload day (2026-08-03), cross-checked against the papers. Independent playbook from Hyperautomation Labs — not affiliated with Stanford.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.