GitHub opens agent workflows that run inside Actions
GitHub has made its open-source Agentic Workflows project generally available. Teams describe recurring repository work in natural language and compile it into workflows for GitHub Actions. Coding agents begin with read-only access, while proposed changes undergo validation and threat checks. The release expands self-hosted execution, model choices and spending forecasts.
Artificial Intelligence··Evening
Natural-language work runs through Actions
GitHub made Agentic Workflows, an open-source project from its GitHub Next development group, generally available on October 28. The project runs coding agents within GitHub Actions, its automation service, for jobs such as sorting issues, investigating failed integration checks, updating documentation and writing tests. A plain-language task written in Markdown becomes a standard Actions workflow in YAML, the service’s structured configuration format. Existing runner groups and organizational policies remain applicable.[1], [2]
Proposed changes pass through separate checks
Agents start with read-only permissions and operate in an isolated container behind the Agent Workflow Firewall, the project’s network-control layer. Access to GitHub content follows integrity filters. The safe-output process validates what the agent produces, and a separate threat-detection job examines proposed changes before application. Each workflow can use Copilot, GitHub’s coding assistant, or another agent chosen by its author. The default Copilot command-line engine requires a Copilot license.[1]
Self-hosted execution and budget forecasts expand
The release adds support for organizations running workflows on their own infrastructure, including Docker containers running inside another container. Spending forecasts leave experimental status and include both the main agent’s use and the separate threat-detection job, with monthly ranges. Model providers can be selected for individual workflows, and routing can use task type and difficulty. New templates accompany the release. Prebuilt examples cover triage, dependency management, code review and reporting.[1]
