Notes on Working with Coding Agents

A solo developer described running up to 5 parallel Claude coding sessions, each in its own git worktree and tackling one GitHub issue at a time, to build a side project over a 2-week holiday. To limit agent access, the developer created a separate non-admin `dev` user on the local machine so Claude sessions cannot reach personal files, and reviews each pull request individually on GitHub. The workflow treats GitHub issues as prompts and documentation, with each session starting from the prompt "Tackle issue #<issue-number> and open a PR.

home https://adri0.com/ Notes on working with coding agents During my last holiday I spent a good chunk of its 2 weeks working on a new side https://github.com/adri0/scadustats/ project https://scadustats.com/ more about it in a future post . A lot has changed with coding agents recently, and it has been a while since I’ve had to push their limits on a relatively complex problem outside my day job. With better models and harnesses, coding agents now seem a lot more competent. This means that now I can unleash Claude s to aggressively work on many ideas in parallel, while I spend more of my time designing the product, running experiments and managing the process - which required me to tweak my solo dev workflow. First requirement: I’m used to keeping a backlog using GitHub issues, even when working alone. I’d like Claude to act like a teammate: collaborating, refining issues and sourcing coding tasks from there. This way the project’s context isn’t trapped only in CLAUDE.md and READMEs. Second requirement: It would be nice to have multiple Claude sessions running in parallel, each solving an issue. Third requirement: I wanted to keep up with understanding the codebase as it evolves, but also learn new patterns and Python tricks suggested by the agents. In addition, I wanted to be able to vet occasional mistakes and bad choices early on, before they get the chance to snowball and there were a few . I’m open-minded about Claude’s approach but reserve the right to shoot down anything that doesn’t smell right. Fourth requirement: I wanted to protect myself from unintended harmful agent actions. I don’t like the idea of my private data ending up in Anthropic’s servers, or worse, sent to some unknown third-party due to malicious injected prompts or any agentic creativity burst. In the end I settled on a workflow that I’m relatively happy with. It boils down to the following: 1. Creating GitHub issues describing individual tasks Issues should contain detailed enough information so that they can be picked up in an individual Claude session and result in a PR. I treat issues almost as a prompt, but it also works as documentation about past product decisions. 2. Creating multiple Claude sessions for tackling problems in parallel Every session almost always starts with the same prompt: “ Tackle issue <issue-number and open a PR. ” I keep up to 5 sessions running at the same time, each on its own terminal. Technically, I could create as many as needed, but it can get hard to keep track of what’s going on. Each session runs in a separate git worktree https://git-scm.com/docs/git-worktree and tackles 1 issue at a time. I envision that the generalisation of this would be having a dedicated agent coordinating the multiple coding sessions, theoretically being able to tackle all open issues at the same time and exhausting my tokens fast . But at the moment it feels overkill. 3. Reviewing open PRs on GitHub I review each open PR individually, just like I’d do if working in a team. Whenever I spot something to be changed or would like clarification I simply leave a comment. Then, I go back to the session that generated the PR and ask Claude to address the open comments. Since Claude uses GitHub CLI gh logged in with my account, its replies show as coming from my own user, looking like I’m talking to myself. 4. Taking precautions against agentic shenanigans I don’t feel very comfortable leaving coding agents on auto mode https://code.claude.com/docs/en/permission-modes eliminate-prompts-with-auto-mode running shell commands while having access to all my personal files. Thus, instead of disabling auto mode and reviewing every command manually, I opted for limiting Claude’s access perimeter: Run on its own user - I created a fresh non-admin dev user on my local machine and only allow Claude sessions to run from this user. This means that Claude doesn’t have access to any file stored in my own user, nor will it have admin access to change anything outside this user space. I still do the work from my main user. I only impersonate this user in a terminal for starting Claude sessions using sudo su - dev https://man7.org/linux/man-pages/man1/su.1.html . myuser ~ sudo su - dev dev ~ cd code/repo Here it is logged in as the dev user dev ~ claude Share a code directory - In addition, the code repos are stored in a shared directory between my main user and the dev user, for the cases where I’d like to inspect the code state or make ad-hoc changes. A drawback from this setup is that I lose Claude integration with my IDE open file awareness, selection awareness, etc. because now Claude and the IDE are running under different users still looking for an alternative . But I can live with that. I don’t feel those features are super essential anyway. Use a fine-grained GitHub token - The last precaution is related to git and GitHub access. I don’t want Claude to have access to my entire GitHub account especially private repos . Thus, for Claude sessions, I log in to GitHub CLI using a fine-grained GitHub personal access token https://docs.github.com/en/authentication/keeping-your-account-and-data-secure/managing-your-personal-access-tokens fine-grained-personal-access-tokens , which is configured to allow access only to the relevant repos, and can write issues, pull requests and comments only in them. Closing remarks This way of working is by no means perfect and I’m sure it will change over time. However, it feels very productive, safe and reasonable. All the paper trail of prompts as issues and pull requests with comments natively tracked on GitHub, mimicking teamwork, feels quite nice. Claude being fairly isolated in its own user, with contained tool access, gives me extra peace of mind.