The concept
End to end means four stages, with you approving each:
- Plan: Claude Code reads the folder and proposes an approach, without editing anything (Plan mode).
- Build: it writes the code.
- Test: it writes tests and runs them until they pass.
- Run: you run the agent on real input and check the output.
What gets built is a small Agent SDK program (Days 72–73). What builds it is Claude Code (Days 69–74).
Walkthrough: an expense-receipt checker
1. Set up a folder, receipt-checker, with:
- policy.md (illustrative rules): hotel up to $180 a night; alcohol not reimbursable; submit within 30 days.
- receipts/: 4 short text receipts you make up. No real personal data.
2. Start in Plan mode: claude --permission-mode plan
Prompt: "Build a Python agent with the Claude Agent SDK. It reads each file in receipts/, checks it against policy.md, and returns file, amount, verdict (ok or flag) and reason. The agent gets read-only tools; the script writes report.csv. Add tests using the 4 samples. Plan first."
3. Review the plan like a product owner. Is anything missing? Say: "Also flag receipts with no date." Approve.
4. Claude builds: agent.py, a tests file, a short README. It runs the tests, reads failures and fixes them. You approve commands as it goes (Day 74).
5. Run it: python agent.py. Illustrative result:
- Hotel, 2 nights, $424 → flag: $212 a night, over $180.
- Taxi, $31 → ok.
- Team dinner, $140 with a wine line → flag: alcohol.
- Printer ink, dated 45 days ago → flag: late.
6. Ask Claude to commit with a clear message.
Two choices that keep it safe: the agent can read but not edit or delete, and it only flags. A person in finance decides.
Where you see it at work
- Banking and insurance: pre-checks on staff expenses and invoices.
- Healthcare: claims or referral forms checked for missing fields.
- Product owners: this plan → build → test → run cycle is a user story, start to finish.
Try it today (10–15 min)
1. Make the folder, policy.md and 4 fake receipts.
2. Install the SDK and set your API key: https://code.claude.com/docs/en/agent-sdk/quickstart
3. Start Claude Code in Plan mode and paste the prompt above: https://code.claude.com/docs/en/permission-modes
4. Approve the plan, let it test, run python agent.py, and check every verdict by hand.
In one sentence
Describe the agent, review the plan, and let Claude Code build, test and run it, while you keep the approvals and the final decisions.
What small checker would you build first at your job? 👇