GitHub launches Agentic Workflows to automate tasks with AI agents

GitHub launches Agentic Workflows

GitHub has launched a new tool called Agentic Workflows. It lets developers create AI agents that automatically perform tasks based on events in a repository, like a new issue being created.

The tool is now available in preview. It uses GitHub Actions as its infrastructure to run agents across millions of repositories.

How Agentic Workflows work

The system is designed for tasks that are repetitive but require subjective judgment, which traditional CI/CD tools can't handle. Eddie Aftandilian, a principal researcher at GitHub Next, explained these are tasks where you "always want an agent watching the events in the repo."

It is not meant to replace existing CI/CD pipelines. Instead, the team calls it an augmentation that adds "continuous AI" capabilities to the development loop.

You can describe a task in plain English, and the tool will generate the detailed workflow steps. Microsoft researcher Peli de Halleux calls this "agentic authoring."

"The barrier to entry is basically all the way to almost zero," de Halleux said. The tool creates two files: a Markdown file describing the workflow and a YAML file for GitHub Actions.

Key use cases for developers

The most basic use is generating a daily status report summarizing recent issues and pull requests. The GitHub Next team provided several other example workflows.

• Continuous triage: Automatically summarize, label, and route new issues.
• Continuous documentation: Keep READMEs and docs aligned with code changes.
• Continuous code simplification: Identify improvements and open pull requests.
• Continuous test improvement: Assess test coverage and add high-value tests.
• Continuous quality hygiene: Investigate CI failures and propose fixes.
• Continuous reporting: Create regular reports on repository health and trends.

While in preview, Agentic Workflows already supports three major coding agents: Claude Code, OpenAI Codex, and GitHub Copilot.

A built-in safety architecture

Permissions are set to read-only by default. An agent can comment or join discussions, but any write operation is deferred and run as a separate job after validation.

"You want the thing that these agents do to be guardrailled, validated, revalidated," de Halleux said. This is critical for giving developers confidence in the system's stability.

The team built on GitHub Actions' security and added a SafeOutputs subsystem. This is a set of trusted components whose outputs are run through deterministic filters to enforce policies.

An additional Agent Workflow Firewall can limit what resources the agents have access to within the repository.

Start with low-risk improvements

The team recommends developers begin with low-risk outputs. This includes having agents generate comments, drafts, and reports before progressing to creating pull requests.

The initial focus should be on improving existing code rather than building new features. Over time, as agents and developer intuition improve, more ambitious use cases will likely emerge.

The official documentation stresses this tool extends automation to subjective tasks traditional CI/CD can't express. The goal is to reduce daily toil for developers and, ideally, result in higher-quality code.