While most of the AI conversation focuses on code generation, GitHub is moving one step further - toward autonomous repository management. It has introduced Agentic Workflows – a feature that allows AI agents to monitor repository events and execute the necessary tasks in repositories based on the triggers or schedules.
The project was initiated by the GitHub Next team, the same team behind Copilot, as a research project exploring what repository automation should look like in the era of AI agents. The result is a model in which the developer describes the desired result in plain Markdown rather than complex YAML, adds the file as a workflow to the repository, and the coding agent handles the execution via GitHub Actions. It carries out repository tasks like triaging issues, reviewing pull requests, investigating CI failures, updating documentation, and more.
GitHub recommends using agentic workflows for any repetitive tasks that can be described in words: from general goals like "improving software" to very specific requirements, such as checking the documentation style for a specific audience. They highlight the following main use cases: Daily project status reports (issues, PRs, discussions) Automatic sorting and tagging of new issues Analyzing test coverage and adding missing ones Keeping documentation up-to-date Checking for best practices with suggested fixes Investigating CI crashes with suggested fixes Instead of hardcoding steps, now it is possible to set a goal, for example, “generate a daily project status report” or “analyze new issues and assign labels”. It is sufficient to just describe the task in a couple of sentences and the system translates them into a step-by-step workflow and selects steps while staying within defined constraints. The focus shifts from writing the perfect workflow script to clearly describing the intended result. Security is central to the design. GitHub Agentic Workflows run like standard workflows in GitHub Actions, but they employ multi-layered protection against unintended actions and prompt-based attacks. By default, workflows are read-only. Any write operations must go through pre-approved sanitized “safe outputs." Additionally, isolated environments, permitted tool lists, and network restrictions are used. This model contrasts with running CLI agents (like Copilot or Claude) directly in YAML workflows, where broader permissions are typically granted. Here, permissions are constrained first, and expanded only when explicitly allowed.
The GitHub Next team calls this approach Continuous AI - the integration of AI into SDLC, similar to CI/CD practices. They highlight that this approach isn't a replacement for existing CI/CD processes, but a complement for existing YAML workflows, enabling the automation of more subjective and repetitive tasks where evaluation, analysis, or contextual understanding are required, but traditional CI/CD pipelines struggle to define rules. Thus, CI/CD remains responsible for the deterministic processes of building, testing, and deploying, while GitHub Agentic Workflows extends automation to include tasks that require reasoning and analysis of code or repository artifacts.
Different code agent engines are already supported, among them are Copilot, Claude Code, or OpenAI Codex. And here is the pricing reality - running an agentic workflow incurs a combination of costs including GitHub Actions compute time and LLM tokens usage. In addition to hosted runner compute costs, GitHub has recently introduced a $0.002 per-minute platform fee. At the same time, it reduced GitHub-hosted runner prices by up to ~39%, offsetting the change for most customers. They also wanted to add this platform fee even on self-hosted runners, but postponed this decision for unknown terms.
LLM costs depend on which agent you choose. GitHub Agentic Workflows defaults to using Copilot as the coding agent. You must have an active Copilot subscription to use it ($19/mo for individuals/business, $39/mo for Enterprise). Most workflow usage is included into that subscription and you generally don’t pay extra, and costs are predictable. Copilot workflows use tokens internally, but for typical repository tasks, your subscription covers it. For many small-to-medium teams using Copilot already, the incremental cost is near zero beyond the subscription.
Other agents follow separate rules, as described in the GitHub documentation. Using Claude or Codex means paying those providers' standard API rates with billed directly to your Anthropic or OpenAI account via API keys stored in Actions secrets. Costs can vary depending on how much the agent works, so they are less predictable than Copilot. The total cost depends on how long the agent runs, how many files it edits, and how many times it requires input. Typical workflows like daily triage, weekly reports, and PR reviews cost about $20–50/month. Either way, these tools can save you hours of repetitive work, making the investment trivial for most teams.
GitHub Agentic Workflows are open source and are being developed in partnership with Microsoft Research and Azure Core Upstream. The feature is now available in technical preview via the gh-aw GitHub CLI extension and the main repo at github.com/github/gh-aw, with documentation and a template gallery already published. Developers are invited to test the new repository automation system, share feedback, and example workflows in community discussions and the #agentic-workflows channel on the GitHub Next Discord.
As GitHub's COO Kyle Daigle mentions in his linked post, announcing the technical preview, Github Agentic Workflows are already being tested in large open-source and enterprise projects. Specifically, Home Assistant has already been analyzing thousands of issues with it, surfacing patterns no one person could track manually. Cloud Native Computing Foundation (CNCF) is transforming documentation and reporting, and consequently their culture. Carvana is running it across complex systems and multiple repos. As the tool runs on top of GitHub Actions, the infrastructure is already ready to scale to millions of repositories.