Guides to running coding agents
Longer pieces on running several coding agents, what they cost, and how to keep the work organized. We use these tools every day and write down what we learn.
How to run several coding agents at once
Give each agent its own git worktree and branch so their edits land in separate folders. Then plan for what a worktree leaves shared: dev server ports, the local database, chat history and your review queue. On one laptop that works for a few agents at a time, and past that the bookkeeping is the hard part.
ReadWhat coding agents cost: Claude Code, Codex, Cursor and API usage
The entry plans that include Claude Code (Claude Pro) and Codex (ChatGPT Plus) cost $20 a month, and Cursor's first plan with cloud agents and other vendors' models costs $30. Heavy users move to $100 or $200 plans. An API key has no monthly cap: Anthropic reports that its enterprise Claude Code users average about $13 per developer per active day.
ReadWhat is a coding agent workspace?
A coding agent workspace is an app that sits above your coding agents and keeps track of the work: which project a task belongs to, which agent runs it, where its conversation lives and whether its result has been reviewed. The agents keep doing the coding. The workspace gives them, and you, one place to organize it.
ReadLoop engineering: let loops prompt your coding agents
Loop engineering means you design the system that prompts your coding agents instead of prompting them yourself, turn by turn. A loop that can run without you needs six parts: a trigger, a written definition of done, a worker in its own worktree, a separate checker, a record of progress kept outside the conversation, and a way to stop or hand the work back to you.
Read
