The Agent Harness Path

How to work a session

1. Read the lesson (20–40 min). The theory-in-depth section is the core; the state-of-the-art table tells you what the industry currently does about it.
2. Run the notebook (30–60 min): uv run marimo edit sessions/sNN-slug/toy.py. Every experiment has a Predict first prompt — write the prediction down before running the cell, then compare what you observed with what you wrote. Prediction misses are where the learning is; a prediction you didn't write down is one you'll retroactively fix.
3. Do the exercises. Attempt cells come before solution cells. The attempt is the rep; the solution is the spotter, not the lift. Running every cell demonstrates execution; an empty attempt skeleton or a score against reference labels is not your independent work.
4. Self-check (5–10 min). Foldable questions at the end of the lesson. Attempt each one before revealing it, then record the correction in your own words; a correct selection does not certify mastery. Nobody is grading you.
5. (Optional) hard path. After the notebook: labs/README.md. Before a lab, read the session's companion.md: it maps the notebook's cafe/ module to the café host's code, which helps when the toy feels too small and the lab feels too sudden. Replay committed cassettes (uv run python labs/run.py --session sNN --replay) or call any OpenAI-compatible endpoint (--live).

That is three required activities per session (read, notebook, self-check): 36 across S01–S12, not twelve automatic completions. The authored hints and self-checks live in the lesson and work without any model or tutor. If you study with an AI tutor in VS Code or Cursor, set it up with docs/COMPANION.md: it explains and reviews, and your predictions and attempts stay yours.

Setup is uv sync --frozen once (the repository README), then uv run python -m cafe.doctor checks the offline stub without keys. Course logic is Python standard library only; marimo is the runtime.

Run against your own model

Notebooks run on the deterministic offline stub by default: no network, no key, no environment file. To run them against your own OpenAI-compatible endpoint, use a dedicated shell and set all three course variables explicitly:

export COURSE_MODE=live
export CAFE_BASE_URL=http://127.0.0.1:11434/v1   # Ollama, LM Studio, anything OpenAI-compatible
export CAFE_API_KEY=ollama                       # any non-empty string for a local server
export CAFE_MODEL=qwen2.5:14b-instruct
uv run python -m cafe.doctor                     # one live call; proves your endpoint

CAFE_* takes precedence over OPENAI_*, with a fallback for each variable; setting all three means a missing course variable cannot select an unrelated credential or endpoint. No key is ever committed, printed, or logged. CI sets stub mode and blocks sockets. These variables configure only the notebooks: an editor tutor's model is set in the editor, and the hard path's --live reads its own OPENAI_* variables (labs/README.md).

Model size matters. Most sessions are better with a mediocre model: bad output is exactly what S02, S07 and S10 measure and repair. S04 (structured generation) and S12 (judge calibration) need a mid-size instruct model with real tool-calling support.

The sessions

S13 and S14 invert the pattern: no notebook by design, and an unaided protocol on the artifact you choose. Read each lesson's prerequisites before deciding to perform it. Their working cards sit beside the lessons: audit-card.md for S13; cold-run-card.md, pilot-page-card.md and pilot-card.md for S14. During unaided work, close the assistant chat and return to review only after the sitting ends and the evidence is saved. Deferring either protocol does not block S01–S12, and completing S01–S12 requires neither them nor labs/. Record planned and performed activities separately; no software test manufactures a rebuild, an elapsed week, a consenting person, a publication or a learning result.

Per-session videos were removed in September 2026: they lagged the lessons and cost more to re-record than they taught. The learning happens in the notebook (S01–S12) or the protocol (S13/S14). The single course overview above stays as an optional preview of the arc.

Core route and post-core overlay

• Core harness route (start here): work through S01–S12, then optionally use S13/S14 and labs/. If you are new to harnesses, complete S01–S12 before using the overlay.
• Calibrated six-week post-core overlay: use the study-plan overlay to schedule authoritative external work from CS336, DeepLearning.AI RLHF, DeepLearning.AI vLLM, and the optional Anthropic API course, with explicit evidence to bank. The overlay is not part of the 14-session core and contains no copied external-course materials. It requires either completed core/hard-path work or equivalent harness experience, plus an inspectable agent/eval artifact with a green, banked baseline.

The model call is the easy part; the harness is the product. The overlay supports two complementary learning emphases: harness/evals/inference-systems depth and model-layer fundamentals depth.

Your work and your feedback

Keep your edited notebooks, predictions and notes in a local, nonsynced folder outside this checkout: a clean clone, branch switch or version change replaces the course files. To report friction, study/FEEDBACK.md gives one useful question per session and says what to include and what to leave out. The course sends no telemetry.

The pedagogical commitments

• One toy domain, never the artifact. Every notebook grows the same neighbourhood-café counter assistant — orders, tickets, allergens, the till — chosen so the mechanics are identical to production and the stakes are not.
• Predict first, always. Written predictions before every run; the suite of habits this builds is the actual curriculum.
• Failure is on the syllabus. Each toy contains deliberately broken variants (labeled) so you meet the failure modes cheaply, where the fix is one cell away.
• State of the art, dated. Each lesson's SOTA table states its review date and every row carries a source. Check evolving specifications and distinguish preprint findings from general guarantees.

License

Apache-2.0 for notebooks, labs Python, and tooling; CC BY 4.0 for lessons, the course overview video, docs, and lab protocols. Cited sources remain their authors'. Google branding in the course overview is Google's. See LICENSE in the repository root.