# Keep a workflow's instructions current

A feedback-report workflow depended on two official Codex pages. dot was given a small continuing job: keep the relevant instructions aligned with those pages and the work actually completed. The useful result was a changed judgment, not another copy of the report.

[Read the saved responsibility](responsibility.md), then inspect [the two recorded checks](observations.json).

## What happened on 24 October 2026

**First check, 10:27 UTC.** Both official pages opened. Their relevant CLI and sign-in guidance supported the workflow's intended route. The package contained instructions and code, but its execution records had not yet arrived in the checked workspace. The saved result therefore said “implementation present; actual run not established here.” It retained the source notes and package state for the next check.

**Later check, 10:37 UTC.** The work owner reported that real execution records and selected outputs had arrived. dot resumed the saved responsibility, opened the same two sources and compared the new work with the baseline. The facts the workflow depended on had not changed. The original package files were unchanged too. What changed was the evidence: two successful, manually started CLI calls now had matching input and output records. A separate duplicate-ID input had been rejected before an AI call.

The next action was to update the workflow's evidence and inspect its existing reports, not rerun either model call. Browser acceptance had not been performed in the execution environment; the continuation retained that limit.

- [First report: 18 fictional feedback items](../feedback-workflow/recorded-run/first/report.html)
- [Later batch: 12 fictional feedback items](../feedback-workflow/recorded-run/changed/report.html)
- [Selected execution record](../feedback-workflow/recorded-run/execution.json)

This was a real manual continuation after new work evidence arrived. It does not demonstrate scheduled or unattended execution. Both source fetches succeeded, so the unavailable-source rule was specified but not exercised. Two synthetic batches do not establish long-term reliability or ordinary-user success.

## Give your Agent a similarly small job

Choose something you already care about: a few public event pages, a course's requirements or the official references behind a project. Tell your Agent what it should look after, which sources it may use, which changes matter, and where to keep the result. Ask it to save the responsibility and its last known state, then confirm where they were saved.

A short brief can be enough: “Look after these event pages for me. Update my saved note when the date, registration or eligibility changes in a way that affects my plans. Keep the last confirmed information if a page cannot be opened, and say what is unconfirmed. Show me where you saved the responsibility and how I can pause it.”

Agree on when it should continue. In this example, the owner supplied the next signal manually. A saved brief alone does not create a background task. For a schedule, use an environment that actually supports it and verify the task was created before relying on it.

## Sources checked

[Non-interactive mode](https://learn.chatgpt.com/docs/non-interactive-mode) documents CLI input and output options used by the workflow. [Authentication](https://learn.chatgpt.com/docs/auth) distinguishes subscription sign-in from separately billed API-key use. Both were read during each check on 24 October 2026. These pages support interface claims; the execution records support what happened in this particular run.