CoursesFree mini course37 min of video

Build an Autonomous Claude Code Agent

Claude Code can fix an issue; the hard part is letting it run while you work on something else. Most teams either watch every step or let it loose with no checks, then stop trusting it after one bad pull request. Live, I give Claude Code a real GitHub issue and run it headless in its own worktree, with hooks and tests, until it opens a pull request that passed them.

Watch the whole sessionAbout the session

The lessons, in order

Each lesson is one idea from the live build, cut to a few minutes, with chapters and the full transcript.

  1. 1

    A Software Factory Is Not a Dark Factory

    A software factory automates what can be automated and sends only what needs a human to review. Bug fixes and dependency updates, light and dark factories, and why the process, not the agent, is what you automate.

    Out
  2. 2

    Boundary and Context Are Yours

    Every company has its own immunity, so every company sets its own boundary for what an agent may do. Then context: the product domain and the technical standards no model knows until you write them down.

    Out
  3. 3

    Skills Stop the Agent Improvising

    A frontier model can figure out your process eight times in ten, but it pays for the research on every run and ten engineers reinvent it ten ways. A skill turns your team's loop into the steps the agent follows.

    Out
  4. 4

    Verify Outside the Agent

    Most teams have execution covered and lack verification. The agent that read the issue will always say its code is verified. Holdout tests belong in the plan, and delivery needs pull requests small enough to review.

    Out
  5. 5

    The Repo Carries the Harness

    A small bill-splitting repo whose .claude directory holds the rules, a fix-issue skill, a reviewer subagent with its own context, hooks and a deny list. Everything the agent knows and may do is checked in for every teammate.

    Out
  6. 6

    Run Claude Headless on an Issue

    One script runs Claude headless with the fix-issue skill on GitHub issue 178, in its own worktree, and opens a pull request. Why a worktree per issue, and how the same run moves to CI or a sandbox.

    Out
  7. 7

    Hooks Refuse Before the Run

    Asked to delete the database and reset the branch, the agent tried both and a hook refused each one before it ran. Autonomous pull requests need gates you trust, the same way a support bot is grounded and evaluated on every release.

    Out
  8. 8

    The Stop Hook Sends Work Back

    The agent wrote a test and said done without running it. The stop hook ran the test, got red and handed the work back; the agent fixed the code, not the test, and only stopped on green. Then a draft pull request and a reviewer subagent.

    Out
  9. 9

    What It Takes to Automate a PR

    A startup with zero users can prompt and ship. A company with revenue needs rules, a skill playbook, hooks, verification gates and holdout tests, so the pull request reads like an executive summary an engineer can merge.

    Out

What the course covers

Engineers who want Claude Code to finish an issue while they work on something else.

Run Claude Code headless on a real issue
Turn a GitHub issue into a task it runs in its own worktree, without you.
Gate its work with hooks and tests
Block unsafe commands and fail the run before a bad change reaches review.
Hand back a pull request worth reviewing
Open the pull request only when every check passes; an engineer merges.

More free mini courses

Every Lightning session becomes one, in the order the Software Factory gets built.

Join a free Live LabSee the buildcamp