Lightning sessionsRecording60 min

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 recording

The whole session is free on Maven. Sign up with your email and the recording opens.

Watch the recording on Maven

Where it sits in a Software Factory

A coding agent is only as reliable as the system around it. This session builds Execution and Verification and Delivery, the layers marked below.

Boundary
Sets what the agent may touch: which files, which commands, which network.
What it stops: An agent editing files outside its task, or running a command it should never run.
Context
Routes the architecture, past decisions and team rules a task needs to the agent working on it.
What it stops: Code that compiles and ignores how your codebase works.
Skills
Keeps engineering know-how in reusable skills, written once and used by every agent.
What it stops: The same instructions pasted into every prompt, drifting apart.
ExecutionThis session
Runs each task in its own Git worktree, inside the boundary.
What it stops: Two tasks colliding, or an agent changing the branch you are working on.
VerificationThis session
Checks the work with static gates, holdout tests and a Doer and Tester loop that sends failures back with the reason.
What it stops: Code that looks right reaching your review.
DeliveryThis session
Opens the pull request only when every check passes, for an engineer to merge.
What it stops: A bad pull request that costs the team its trust in the agent.

Why this matters

Claude Code can fix an issue; the hard part is letting it run while you do something else. Watching every step defeats the purpose, and letting it loose with no checks ends in a pull request nobody trusts. Running it headless in its own worktree with hooks and tests gives you work you can review, not babysit.

What you will learn

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.

The ideas behind it

  • Headless on a real issue

    Claude Code takes a GitHub issue and works without a chat window.

  • Hooks and tests as gates

    Checks that run on every change and stop work that fails them.

  • A pull request worth reviewing

    The run ends only when the checks pass, with the reasoning in the description.

The series, in order

Each session builds on the one before. Start anywhere: every recording stays on Maven.

  4. Build an Autonomous Claude Code AgentExecution, Verification, Delivery
    This session

Before you come

For
Engineers who want Claude Code to finish an issue while they work on something else.
You need
Claude Code, Git and a GitHub repository.
Price
Free on Maven. You sign up with your email.
Can't make it live?
The recording is on Maven.

Questions

Is it safe to leave it running?

Its worktree, tool list and hooks limit what it can do, and nothing merges without you.

Does it need a paid Claude plan?

It needs Claude Code with an API key or a plan that includes it.

Is it free?

Yes. It is a free Lightning session on Maven, and signing up gets you the recording.

After this session

The Live Labs build one agent system piece by piece, free, with time for your questions. The buildcamp builds all of it with you over four weeks. See the Live Labs or the buildcamp.

More Lightning sessions