When your agent stops and you have to ask it to finish

Definition

A nudge is a message you send to a coding agent so that it resumes an unfinished task you had already authorized: “finish”, “commit”, “check CI”. The agent had the information and the permissions it needed to continue, and it stopped.

Avoidable check-ins

An avoidable check-in is the agent asking you to authorize an action you already authorized. For example, it asks “Want me to commit this?” when your instructions included the commit. Your “yes” tells it to continue: that answer is a nudge.

What is not a nudge

Four situations come close without being one.

• A real decision: the agent asks a question that is yours to answer, such as which of two designs to keep.
• A session waiting: the agent has finished or is waiting for your answer while you are busy elsewhere.
• A project switch: you leave one repository for another to answer an agent.
• Approving a risky or irreversible action, such as deleting data or publishing. Those confirmations should stay. Humanload will count them as decisions.

Loops and written instructions

A loop restarts the agent for you. In its simplest form, Geoffrey Huntley’s “Ralph” technique is a Bash loop that feeds the same prompt to the agent again and again. Huntley presents it mainly for starting new projects, and writes that he would not use it in an existing codebase. Business Insider describes /goal, a command that tells Claude Code or Codex to keep working until a task is complete, and reports that cost is the main concern raised about loops.

A written instruction targets the cause. You put it in the instruction file your tool loads for the project, instead of repeating it in every session. For example: “Commit to a branch without asking when the task includes it. Ask before anything risky or irreversible, and when a decision is mine to make.” Claude Code reads CLAUDE.md, or AGENTS.md when there is no CLAUDE.md. Its documentation says these files are context, not enforced configuration, and that short, specific instructions are followed more consistently.

Huntley tunes his own loop with what he calls signs: written instructions placed where the agent keeps going wrong.

Claude Code’s auto mode removes routine permission prompts: a separate model reviews actions before they run. Its documentation states that this reduces prompts but does not guarantee safety.

What Humanload will count

Humanload will count nudges alongside decisions, project switches and the sessions waiting for you. If you turn on your retro, it will suggest an instruction for your AGENTS.md or CLAUDE.md when the same nudge keeps coming back. With your agreement, it will open a pull request, which you merge or close. Anyone with access to the repository can see that pull request. Humanload will then keep counting nudges of that kind.

Humanload is in development, and this page describes what it will do. It will not stop nudges for you, and it will not change your files outside that pull request.

Sources

• Geoffrey Huntley, “Ralph Wiggum as a ‘software engineer’”, July 14, 2025.
• Business Insider, “Forget prompt engineering: ‘Loop engineering’ is all the rage now”, June 20, 2026.
• Claude Code documentation, “How Claude remembers your project”, read October 9, 2026.
• Claude Code documentation, “Choose a permission mode”, read October 9, 2026.
• AGENTS.md, an open format for coding agents, read October 9, 2026.

The waitlist

Join the waitlist to get an email when Humanload launches.