The Pulse
GitHub Moves Agentic Workflows Into Actions Public Preview
GitHub has moved Agentic Workflows into public preview, allowing coding agents to run repository automations inside GitHub Actions. Developers describe tasks in Markdown, while the gh aw extension compiles them into standard Actions workflo
AI.info Team ·
Two files are enough to turn a plain-language repository instruction into a running GitHub Actions workflow: a Markdown file describing the job and a compiled lock file that Actions executes. GitHub is now making that model available in public preview through GitHub Agentic Workflows, bringing coding agents into the automation system used for testing, deployment and repository maintenance.
GitHub announced the public preview on June 11, 2026, after introducing the project as a technical preview in February. The company says the workflows can handle tasks such as issue triage, continuous integration failure analysis and documentation updates, with the work performed by coding agents inside GitHub Actions.
The current GitHub documentation describes the feature as a way to define repository automation in Markdown and select the coding agent that runs it. The gh aw extension compiles that configuration into a standard GitHub Actions workflow.
Markdown instructions become Actions jobs
Agentic Workflows separate the task description from the execution machinery. Developers write the intended outcome in natural language, add configuration for triggers, permissions, tools and allowed outputs, then use the GitHub CLI extension to generate the workflow that runs in Actions.
GitHub’s earlier technical-preview announcement named Copilot CLI, Claude Code and OpenAI Codex as possible agent engines, depending on configuration. The current documentation also describes using Claude Code, OpenAI Codex or Google Gemini CLI, with the relevant API key stored as a repository secret when required.
GitHub’s own example creates a daily repository status report. The workflow can examine issues, pull requests, discussions, releases and code changes, then publish a report for maintainers. Other examples include summarizing and labeling incoming issues, keeping documentation aligned with code, improving test coverage, investigating failed CI runs and producing recurring repository-health reports.
Read-only defaults limit what agents can change
GitHub is presenting the feature as an extension of existing Actions controls rather than a replacement for deterministic build and release pipelines. Agentic Workflows reuse runner groups and policy constraints, while adding controls intended to keep an agent’s actions bounded.
According to GitHub, agents run with read-only permissions by default. Write operations pass through “safe outputs,” a set of approved operations such as creating a pull request or adding an issue comment. The system also uses a sandboxed container, tool allowlisting, network isolation and a separate threat-detection job that scans proposed changes before they are applied.
Pull requests are not merged automatically. GitHub’s documentation and launch material place a human review step between an agent’s proposed change and its acceptance into a repository.
“Getting an agent to open a pull request was never the hard part. Trusting it enough to merge is.”
May Walter, CTO, Hud.io
GitHub targets repetitive work across repositories
The feature is aimed at work that requires judgment but follows a recurring pattern. A maintainer might ask an agent to identify the most important new issues, a platform team might use it to investigate recurring test failures, or an engineering group might have it open narrowly defined pull requests for refactoring and test improvements.
GitHub’s February announcement described the approach as “Continuous AI,” while stressing that it should sit alongside continuous integration and continuous deployment. Traditional YAML workflows remain responsible for deterministic builds, tests and releases; the agentic layer handles tasks that are difficult to express through fixed rules alone.
Alex Devkar, senior vice president of Engineering and Analytics at Carvana, said the company is using the workflows for engineering work that spans multiple repositories.
“With GitHub Agentic Workflows, we’re able to expand how we apply agents to real engineering work at scale, including changes that span multiple repositories. The flexibility and built-in controls give us confidence to leverage agentic workflows across complex systems at Carvana.”
Alex Devkar, senior vice president, Engineering and Analytics, Carvana
Cost and review remain part of the design
Agentic Workflows consume model capacity and can create billing costs. GitHub’s technical-preview documentation said a default Copilot run typically incurred two premium requests: one for the agent’s work and another for a safe-output guardrail check. Teams can configure the models used and should begin with low-risk outputs such as comments, drafts and reports before allowing pull-request creation.
The public preview leaves the workflow definition itself under source control. GitHub advises teams to treat the Markdown instructions as code, review changes to them, keep their scope narrow and evolve them deliberately.
GitHub is positioning the feature as a way to place agents inside the repository processes developers already inspect through Actions logs, pull requests and policy settings. Its public-preview status also means the feature can change, and the company is asking users to test it through the documentation and quickstart process before adopting it for higher-risk changes.