Start here

This site is a set of working notes on developing with coding agents. It is not a pitch. It describes one setup that works, why each piece is there, and what each piece costs you.

You review, agents write

The premise is simple: you direct and review, the agent writes the code. You still make the decisions that matter (what to build, what the design should be, whether the diff is correct), but you stop typing the implementation yourself.

The practical version of this is a habit, not a tool. Once you commit to it, manual labour starts to read as a smell. If you catch yourself thinking “I will just hand-edit these forty files” or “this is only twenty minutes of clicking”, that is usually a sign you should be writing a script or handing the job to an agent instead. Manual work is not just slow; it drifts the moment the underlying data changes, and it leaves nothing behind that you can rerun.

The honest counterweight is that this only works if you actually review. Agents produce plausible code quickly, and plausible is not the same as correct. Silent partial success is the expensive failure mode: a migration that “completed” while skipping records, a test suite that “passes” because half of it was skipped. Read the diff. Ask for the verification output rather than the claim.

The setup at a glance

• A local machine as the default dev box. These notes assume a Mac, but the shape carries over to any Unix machine. Your laptop is where work happens unless you have a specific reason for it to happen elsewhere.
• Claude Code as the primary agent. Terminal-native, works on the repo in front of it, and configurable per project through a CLAUDE.md file.
• herdr as the agent multiplexer. One agent in one terminal is easy. Four agents across three repos is not, and herdr is what makes that manageable: one window for the whole herd, with sessions that survive closing the terminal.
• A cheap Hetzner VPS as an optional always-on box. Useful when you want long jobs to keep running while your laptop sleeps, or when you want a machine you can reach from a phone. Reached over Tailscale, with public SSH firewalled off entirely.

The VPS is genuinely optional. Plenty of people run this whole workflow on one laptop and never miss it. Add it when you notice yourself waiting on your own machine, not before.

How the guides fit together

Read them in roughly this order:

1. Local (macOS) setup gets your everyday machine into shape: shell, prompt, tmux, package managers, Claude Code, and the small configuration details that otherwise cost you an afternoon.
2. Claude Code workflow covers how to actually drive the agent day to day: the shell aliases worth having, CLAUDE.md conventions, plan mode, and memory.
3. herdr is worth setting up as soon as you find yourself running more than one agent at a time.
4. Skills are reusable instruction packs that teach agents specific workflows. Install them once you have the basics working; they are an amplifier, not a foundation.
5. VPS setup (Hetzner) is the last step, and only if you want an always-on box.

Speed and safety are a real trade

A lot of agent workflows lean on permission-skipping modes so the agent stops asking before every file write and shell command. That is a genuine trade: you get flow, and you give up the checkpoint that would have caught a bad command before it ran. It is reasonable in a scratch repo or a disposable VPS, and much less reasonable on a machine holding credentials or production access. The Claude Code workflow page covers what exactly gets skipped and how to scope it.

What you need

• A machine. A Mac for these guides as written; any Unix machine with a bit of translation.
• A Claude subscription for Claude Code. The agent is the point; everything else here is scaffolding around it.
• A GitHub account with SSH access configured. Agents work best against version control, and several steps in these guides assume gh is authenticated.
• Optionally, a Hetzner account (or any VPS provider running Ubuntu 24.04) and a Tailscale account if you want the always-on box. Both have free or near-free tiers for a single-user setup.

Nothing here requires you to adopt all of it. The local setup on its own is worth the hour it takes. Everything after that is you deciding how much of the herd you want to run.