Phase 10 — Capstone

Part of Harness engineering.

One lesson joins the parts of the track in a single file. The agent has a loop, tools, a permission gate, output limits, a todo list, and a step budget. A test runs it against a real temp repo and checks the repo, not the agent’s words.

Lessons

1. Assemble the agent — One 150-line agent, a scripted model, and a test that proves the parts work together.

Test yourself

1. Why does the test judge the agent by running the repo’s own tests? <details><summary>Answer</summary>The agent can say “fixed” and be wrong. The repo’s test run is an independent check of the real outcome. (<a href=“./01-assemble-the-agent/docs/en.md”>Lesson 1</a>)</details>
2. A command chains an allowed test run with rm -rf .. What does the gate decide, and why? <details><summary>Answer</summary>It decides deny. Deny rules run first, and the rm pattern matches anywhere in the command. The allow rule never gets a vote. (<a href=“./01-assemble-the-agent/docs/en.md”>Lesson 1</a>)</details>
3. Why must trimming drop old turns in pairs? <details><summary>Answer</summary>A tool call and its result must stay together. If one is dropped alone, the model sees an orphan, and the API can reject the history. (<a href=“./01-assemble-the-agent/docs/en.md”>Lesson 1</a>)</details>
4. What does the step budget do when the model loops forever? <details><summary>Answer</summary>The loop stops after the set number of model calls and returns status budget. Every call made so far has its result in the history. (<a href=“./01-assemble-the-agent/docs/en.md”>Lesson 1</a>)</details>
5. What changes when you swap the scripted model for the real SDK? <details><summary>Answer</summary>Only the one function that turns history into content blocks. The loop, tools, gate, limits, and tests stay the same. (<a href=“./01-assemble-the-agent/docs/en.md”>Lesson 1</a>)</details>