How does agentcells keep the AI from going off the rails?

The AI is never allowed to write to GitHub — label changes, pushes, PR creation and comments are all done by the CLI. The next destination must be one of the allowed values in the claude -p --json-schema enum, and back-and-forth is capped at 3 rounds per human reply.

What happens when the AI gets stuck?

It writes the reason and hands the issue back to a human via the ac:needs-human label. Examples given include stopping before review because tests could not be run, or because there was no new commit to send to review.

Why not just tell the AI to ask a human when unsure?

The developer found that a prompt instruction alone leaves compliance up to the AI. Enforcing it in the system makes the reason for stopping explicit when a rule blocks progress.

Large Language ModelsAI Coding AssistantsZenn AI/ML1h agoOct 26, 2026, 01:00 JST

agentcells: AI runs GitHub Issues, but never touches GitHub

Send on LINE

3 Key Points

What happened
The developer behind agentcells, a personal project that splits work into 'cell' units (triage, implement, review), says the biggest win was banning the AI from writing to GitHub. The CLI handles label swaps, pushes, PR creation and comments; the AI only returns a result, with next steps restricted via the --json-schema enum in claude -p and a max_rounds = 3 cap.
Why it matters
Letting the AI pick only from allowed destinations and blocking its own GitHub writes makes the workflow deterministic, so when a rule stops the run the AI reports why (e.g. 'could not run tests, so stopping before review') instead of the developer hoping a prompt instruction is followed.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Get these free every morning

Context & Analysis

The project is a personal rebuild of a workflow the developer already runs at work, and the write-up focuses on what changed when starting from scratch. Two lessons stand out. First, GitHub label search does not immediately reflect a label change, so the same cell ran twice on the same issue; the fix was to read open issues and PRs via GraphQL and filter by label there. Second, running claude -p unattended surfaced process-level risks: repository settings and hooks are not loaded (--setting-sources "" is used, and CLAUDE.md is requested in the prompt instead), and builds or test servers started through Bash can outlive a timed-out claude run because they sit in a separate process group — those leftover processes could read the GitHub token the CLI later prepares for pushing, so the CLI hunts down and stops AI-started processes before any write. For public repositories, the developer limits runs to issues assigned to himself with labels he applied, treats only his own comments as instructions, whitelists readable paths in the sandbox, and has the CLI scan for secrets and changes to .claude/ or .github/workflows/ before pushing. The developer notes some issues remain and are being worked on.

FAQ

How does agentcells keep the AI from going off the rails?: The AI is never allowed to write to GitHub — label changes, pushes, PR creation and comments are all done by the CLI. The next destination must be one of the allowed values in the claude -p --json-schema enum, and back-and-forth is capped at 3 rounds per human reply.
What happens when the AI gets stuck?: It writes the reason and hands the issue back to a human via the ac:needs-human label. Examples given include stopping before review because tests could not be run, or because there was no new commit to send to review.
Why not just tell the AI to ask a human when unsure?: The developer found that a prompt instruction alone leaves compliance up to the AI. Enforcing it in the system makes the reason for stopping explicit when a rule blocks progress.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.