Week 8: the capstone, split, packaged and defended | Agent School | deployed

Portfolio mission. Goes in your own repos, in public. You turn eight weeks of work in one folder into repositories a stranger can understand, run and trust.

Start here if you have work worth showing but it lives in one tangled folder. Skip it if you already have three or more public repos with kept history, a README with reproducible numbers and an eval report.

The three ideas you need

A portfolio is evidence a stranger can verify in ten minutes. A README answers what it does, how well, at what cost and how to run it, in the first screen. Numbers sit next to the command that reproduces them. Field notes: "How do you present numbers people can check?" and "How do you explain your system to someone who will never read the code?"
History is part of the proof. Splitting one journey repo into several should keep each project's real commits. A single commit named "initial" suggests the work was pasted in. Real history shows the eval written before the improvement, the churn caught, the fix. Field note: "What is a decision trace, and why write down the why?"
One agent becomes several only when a measured problem says so. Splitting adds cost, delay and new ways to fail, because agents pass messages and can disagree. It pays when the pieces need different tools, different permissions, or a second opinion. Field notes: "When does one agent become several?" and "What breaks when two systems quietly disagree?"

Who builds

You do not type the code. You write the brief and the prediction; your agent builds and runs the git commands; you read the result. The design choice and the presentation are yours.

The mission

Write BRIEF.md first (what the split must produce, what must never happen: a secret or a squashed history in a public repo). Log choices in DECISIONS.md.

Split your journey repo into at least three repos, for example extraction (Projects 1 and 2), knowledge agent (Project 3), hardened unattended agent (Project 4). Use git filter-repo --path <folder> on a fresh clone, or git subtree split --prefix <folder>. Do not squash. List them in REPOS.md: one Repo: <local path or URL> per line, the main one first, and Split date: YYYY-MM-DD.
Clean each repo. Remove secrets (read the hits of git log -p | grep -iE "key|token|secret"), use sample or synthetic data where the real data belongs to someone else, add a LICENSE and a .env.example.
Write each README with sections ## Results (a table), ## Reproduce (the exact command in a fenced block), ## Cost (per thousand rows, from usage or your provider's equivalent), ## Limits, and a link to the eval report. The main repo's README has a line Reproduce: <command> that prints overall: NN%.
EVAL_REPORT.md in the main repo, sections ## Held-back accuracy, ## Cost curve, ## Knowledge agent scores, ## Churn event, ## Injection test, plus Held-back accuracy: NN%. Every number links to the file or command that produced it; a number you cannot reproduce is deleted.
DESIGN_CHOICE.md, one page: for your capstone, one agent or several? First line under ## Answer is one agent or several; then ## Evidence, ## What would change my mind, and Comparison: ran or Comparison: argued. If you ran one (a planner plus a reviewer against your single agent on the held-back rows), save comparison.csv (config,accuracy,cost_per_1k).
PRESENTATION.md: an eight-minute script under ## Problem, ## Result, ## Cost, ## The failure I am proudest of catching, ## Design choice. Record it if you like.

Deliverable and how it's checked

Three or more public repos with README and history, REPOS.md, EVAL_REPORT.md, DESIGN_CHOICE.md, PRESENTATION.md, BRIEF.md, PREDICTION.md, DECISIONS.md. "Defended" means /agent-school:defend: the examiner asks you to defend the design choice with your own numbers and writes its feedback into COURSE_PROGRESS.md. Nothing blocks you, and there is no outside score: your own reproduced numbers are the evidence. The local check is /agent-school:check week8.

Local check contract: /agent-school:check week8
- `REPOS.md` has >=3 lines `^Repo: \S+` and `^Split date: \d{4}-\d{2}-\d{2}$`
- for each `Repo:` that is a local path: it is a git repo, `git rev-list --count HEAD` >= 10, and the first commit's date is earlier than `Split date` by >=1 day
- for each `Repo:` that is a URL: needs network access; `git ls-remote <url>` succeeds with no credentials (public check); it is then cloned and checked as above
- each repo has `README.md` with `## Results`, `## Reproduce` (followed by a fenced block), `## Cost`, `## Limits`; a `LICENSE` or `^License:` line; `.env.example`; no `.env` in `git ls-files`
- no match of `sk-[A-Za-z0-9_-]{20,}|AKIA[0-9A-Z]{16}|ghp_[A-Za-z0-9]{30,}` in the output of `git log -p --all` run in each repo (120 s limit per repo)
- `EVAL_REPORT.md` has the five `## ` sections above and `^Held-back accuracy: \d+(\.\d+)?%`; every relative markdown link target exists
- reproduce check: fresh clone of the main repo, run its `Reproduce:` command; stdout has `^overall: (\d+(\.\d+)?)%` within 2 points of `Held-back accuracy` in `EVAL_REPORT.md`
- `DESIGN_CHOICE.md` is <=500 words, has `## Answer` whose first non-empty line is `one agent` or `several`, `## Evidence`, `## What would change my mind` (>=1 line), and `^Comparison: (ran|argued)$`; if `ran`, `comparison.csv` has >=2 rows and header `config,accuracy,cost_per_1k`
- `PRESENTATION.md` has the five `## ` headings above and 600 to 1300 words
- `BRIEF.md` four sections; `PREDICTION.md` has `## Predictions` (3 numbered lines) and `## The gap`; `DECISIONS.md` has >=2 entries

Examiner probes:

Defend the design choice: what numbers would flip it?
Which number in your report is weakest, and what would it take to trust it?
What will a stranger fail to understand in your README?
Which decision in DECISIONS.md would you now reverse?

Predict before you build

Write these in PREDICTION.md under ## Predictions first.

How many commits will your oldest project have after the split?
Will a multi-agent version beat your single agent on accuracy, cost, both or neither?
What is the first question a stranger will ask that your README fails to answer?

When you're stuck