GitHub Agentic Workflows is generally available

GitHub Next's open source project lets you write automation in Markdown for tasks such as issue triage, CI failure analysis, documentation updates and test writing, using Copilot or another coding agent inside GitHub Actions. The Markdown compiles to standard Actions YAML, so workflows reuse existing runner groups and policy constraints. By GitHub's account, agents run with read-only permissions by default in a sandboxed container behind a firewall, outputs go through a validation step, and a separate threat detection job scans proposed changes before they are applied.

The general availability release adds support for self-hosted ARC runners with Docker-in-Docker, a gh aw forecast command for AI credit spend, per-workflow model provider choice, model routing, and micro-VM isolation for agents.

GitHub's announcement includes figures from Ben Alton of Accenture's Strategy, Architecture & Innovation team: AI authors 91% of his team's commits, and production errors fell from about 11,000 a day to under 1,000. These are a customer's statements published by GitHub, with no method given.

Why it matters: The safeguards GitHub lists are the ones a team would otherwise build for itself: read-only defaults, a sandbox, validated outputs and a scan before changes land. The page describes them without testing them, so treat them as claims until you have run a workflow against your own repository.