A reliable agent might leave the code alone today

The original Chinese-language cover labels read “Minimal precheck,” “Leave unchanged,” and “No changes needed.”

A reliable agent might leave the code alone today#

On September 11, GitHub Agentic Workflows published a small automation agent called jsweep. Its task is narrow. Every day, it picks one .cjs file from actions/setup/js/, prioritizing files that still contain @ts-nocheck, and performs a small cleanup. The interesting part is its ability to decide that no change is needed today, rather than how it replaces var with const.

Before changing anything, decide whether it is worth doing#

jsweep does not immediately read the whole candidate file. It first starts a file-triage subagent that reads only the first 80 lines. That subagent answers specific questions: what context the file runs in, whether it has @ts-nocheck, whether a corresponding test file exists, and whether the cleanup is worth continuing. If the result is noop, the main agent stops. This step makes a decision before taking action. The agent does not need to fill its context first and then find a justification for doing something.

The September 11 run produced no PR#

The run GitHub published that day finished in under 8 minutes, used 336K tokens, and took only 2 turns. The final result was zero errors, zero warnings, and zero PRs. This was expected. The official page explicitly explains that producing no PR was a designed outcome because triage decided the candidate did not justify a diff. Its safe-outputs configuration supports that decision. create-pull-request uses if-no-changes: ignore, gives PRs a 2-day expiration, and requires them to be drafts. The system does not make creating a PR a condition of success.

Why many agents produce pointless actions#

If a task has just one goal, such as “optimize the code every day,” a model can easily treat making a change as evidence of completion. That looks proactive in a demo. In long-running automation, it creates two problems. First, low-value diffs enter the code review queue. People spend time confirming that the changes were unnecessary. Second, the agent spends more context, tool calls, and model turns on a task that should have ended earlier. jsweep does not solve every cost problem. It still used 336K tokens that day, so this case cannot be described as a low-cost agent. But it clearly separates the decision about whether to continue into its own step.

Noop belongs in the task protocol#

I would make noop part of an agent’s output protocol rather than an exceptional branch. A code maintenance agent might have three possible results. The first is a clear problem that leads to a draft PR. The second is a problem without enough tests, or with too much risk, that leads to a report awaiting human action. The third is nothing worth changing, which returns noop. All three complete the task. They differ in the action they take. This also makes evaluation easier. You can evaluate whether the agent stops when it should, rather than merely count how many changes it successfully submits.

How builders can implement this#

If you are building a long-running coding agent, start with four changes. First, narrow each task. Do not make tidying the entire repository a single run’s goal. Limit the scope to one file, one issue, or a clearly defined module. Next, add a precheck. Read only the minimum information needed for the decision. Its goal is to decide whether the main task is worthwhile, not to solve the problem. Then define noop. Give it explicit return conditions and a log format. Do not treat no changes as a failed run. Finally, constrain the outputs. Automated code changes should default to drafts. High-risk operations need human approval. If nothing changes, do not create an empty PR or a report just for form’s sake.

This is closer to a production concern than a smarter prompt#

One difficulty with production agents is that they run repeatedly. One unnecessary change may seem harmless. Across daily runs, dozens of repositories, and hundreds of tasks, low-value actions become a direct cost in human review time. That is why I care more about an agent’s stopping conditions, scope, and safe outputs than whether every run can generate something that looks complete. jsweep is a small example, but it states these controls clearly. Knowing when to leave code unchanged is part of task design for long-running automation.#