Practice mode · free · unlimited · not certifiable

AI Command-Line Agents

Unscored tasks with hints and instant feedback. Practice results never become credentials.

Short lesson

Working with agents safely · 10 min

Name each agent by role and task, isolate them in worktrees, hand off through files, review every diff, never paste secrets. Tool notes (2026-10): Claude Code (CLAUDE.md, agents folder), Codex CLI and others (AGENTS.md), Gemini CLI (GEMINI.md), Cursor, Copilot — verify names in each tool's docs.

L1Run an agent on a fileterminal

In ~/shop, start a mock agent named builder-readme (profile steady), ask it to update the readme with setup steps, review its diff, and approve it.

L3Approve the good, reject the badterminal

Two agents are available in ~/shop. Start builder-fast (profile swift) and builder-careful (profile steady). Ask both to “add input validation”. Read both diffs. Approve only a correct change and reject a change with bugs.

L4Stop an agent going off trackterminal

Start builder-ratelimit (profile swift) in ~/shop and give it “add rate limit”. Review its proposal: if it is wrong, reject it with a reason, then re-run the task with a careful agent named builder-ratelimit-2 (profile steady) and approve the correct version.