GitHub Agentic Workflows: start with one clear job

Repeated upkeep can have a clear starting point

If you are building an app, try writing down one upkeep job that keeps getting postponed. A short summary of unfinished work can be a useful first experiment. Decide who will read it and what they should learn.

On October 28, GitHub announced the full release of Agentic Workflows. The project, whose code people can inspect and adapt, lets people describe recurring tasks in a Markdown document, a plain text file with simple formatting. It turns that description into work that runs through GitHub Actions, the service that carries out tasks around a code project.

GitHub gives examples such as sorting incoming issues, explaining failed checks, and updating documentation. The difference from a fixed checklist is that an AI system can interpret the information it finds. That also makes reviewing its results part of the job.

Read the announcement: GitHub: Agentic Workflows generally available.

Know what it can reach and what it costs

The release is available on github.com for GitHub Free, Team, and Enterprise accounts. Using Copilot needs a Copilot licence. Other supported AI systems can be used, with their own account and usage requirements. Computer time used by GitHub Actions also needs to be included in your budget. Check the provider's actual bill too; estimated AI costs can differ.

GitHub says agents start with read-only access, meaning they can look at information without directly changing it. Proposed changes pass through separate controls. These protections support careful setup; they do not make every instruction or result correct.

The official starter guide requires a project you can manage, GitHub Actions enabled, an AI account, and a tool for running GitHub instructions from your computer. This is a configured building tool, rather than a switch that runs your business automatically. A guided first setup may be the sensible choice.

Begin with information, then review it

For a first exercise, ask for a daily summary of unresolved items in a practice project. Keep it away from customer information. Give it a short time limit, agree a modest budget and check the actual charges, and decide when it should produce no report.

Review the first few reports against the original items. Can you see what evidence supports each statement? Did it miss something important, repeat yesterday's update, or suggest a next step that belongs to a person? Those observations give you a useful way to improve the instructions.

1. Choose one question the report should answer, such as which items need an owner.
2. Limit the routine to the practice project and the information it needs.
3. Set its schedule, time limit, budget, and named reviewer before enabling it.
4. Compare each result with the source, and pause the routine if the results need correction.

Build the habit before expanding the job

Our view is that managing a useful AI routine is a skill worth practising: explain the job, limit its access, and check the outcome. Once a report helps someone take a better next step, you can consider another small task. Let each added permission have a clear reason.