GitHub Agentic Workflows: Markdown Replaces YAML as AI Enters the CI/CD Loop

On February 13, 2026, GitHub launched Agentic Workflows in technical preview — a fundamentally new approach to repository automation that replaces YAML configuration with plain Markdown and puts AI coding agents at the heart of CI/CD.

What Are Agentic Workflows?

Instead of writing complex YAML workflow files, developers now describe their automation goals in natural language Markdown. Drop a .md file into .github/workflows/, and the gh aw CLI compiles it into a standard GitHub Actions workflow. When triggered, an AI coding agent — GitHub Copilot CLI by default, but also Claude Code or OpenAI Codex — executes the described tasks within a sandboxed Actions runner.

The concept GitHub is calling Continuous AI: the integration of AI agents into the software development lifecycle alongside CI/CD, not replacing existing automation but augmenting it with intelligent decision-making.

What Can They Do?

GitHub has outlined several categories of automation that would be difficult or impossible with traditional YAML workflows:

Continuous Triage — Automatically summarize, label, and route new issues as they come in
Continuous Documentation — Keep READMEs and docs aligned with code changes after every merge
Continuous Code Simplification — Proactively identify refactoring opportunities and open PRs
Continuous Test Improvement — Assess test coverage gaps and generate high-value tests
CI Failure Investigation — Analyze failing builds and propose fixes automatically

Each of these runs as a standard GitHub Actions workflow with familiar triggers: issue events, pull request hooks, cron schedules, manual dispatch, or even comment commands.

Security-First Design

GitHub has clearly anticipated concerns about AI agents running unsupervised in repositories. Agentic Workflows enforce several guardrails by default:

Read-only permissions — Workflows can read repository contents but write operations go through "safe outputs," a sanitized layer for creating issues, PRs, and comments
Sandboxed execution — Agents run in isolated Actions runners with network restrictions
SHA-pinned dependencies — No supply chain surprises
Full auditability — Every action the agent takes is visible in the Actions log, making it inspectable and reviewable

This is a meaningful departure from the "just trust the AI" approach. Every workflow run leaves a clear trail.

The Markdown Authoring Experience

Perhaps the most compelling aspect is the authoring model. A workflow file might look something like:

# Triage New Issues

When a new issue is opened, read the issue body, check for
similar existing issues, add appropriate labels, and leave
a comment summarizing the issue for the team.

The gh aw CLI extension compiles this into the Actions YAML under the hood. Developers can also use AI agents themselves — in VS Code, on github.com, or via their preferred coding agent — to create, edit, and debug these workflow files. It's agentic authoring of agentic workflows.

Open Source and Extensible

The entire system is open source under MIT. Workflows can integrate with the GitHub MCP Server for native access to repos, issues, PRs, and security features. Additional MCP servers can be configured for browser automation, web search, or custom tooling.

The choice of coding agent is also flexible. While Copilot CLI is the default, teams can swap in Claude Code, OpenAI Codex, or other agents — same Markdown format, different execution engine.

Why This Matters

GitHub Agentic Workflows represent a shift in how we think about repository automation. YAML workflows were powerful but rigid — they encoded how to do things step by step. Markdown workflows encode what you want to achieve and delegate the how to an AI agent.

For the millions of repositories already using GitHub Actions, this is a natural evolution. The infrastructure is familiar, the permissions model is proven, and the new layer adds intelligence without sacrificing control.

The technical preview is available now through the gh aw CLI extension. Whether it lives up to the "Continuous AI" vision will depend on how well the agents perform in real-world repositories — but the architecture is sound, and the developer experience is remarkably low-friction.

Sources: