Agentic workflows move from prompt to Actions

The promise of an agent in a repository is easy to describe and hard to operate. Someone still has to define the permissions, runner, budget, evidence, and point at which a proposed change becomes a real change. GitHub Agentic Workflows puts those concerns inside the existing Actions surface.

GitHub says the open-source GitHub Next project is now generally available. A workflow is written as a Markdown file with natural-language instructions and compiled into standard Actions YAML. The documented use cases include issue triage, CI failure analysis, documentation updates, and test writing.

The useful unit is a bounded job

The release is more interesting as an operating model than as a new prompt format. A team can reuse runner groups and policy constraints, start with read-only permissions, and add threat detection and safe-output checks before allowing changes. That makes it possible to test an agent against a completion contract instead of trusting a chat transcript.

Start with the repository task you can review, not the repository task you wish you could delegate.

A sensible first workflow might classify new issues or summarize a failed build without writing to the repository. Define what evidence the output must contain, how much it may cost, who reviews it, and what it is forbidden to touch. Autonomy is easier to expand when the first boundary is explicit.