📚

Core Concepts

Foundational understanding of GitHub Agentic Workflows

✦ Technical Preview⚙ v0.64.0⭐ 4,167 Stars📦 MIT License

What is GitHub Agentic Workflows?

AI-powered automations that understand context, make decisions, and take action — from natural language instructions in Markdown.

Unlike fixed if-then rules, agentic workflows use coding agents (Copilot, Claude Code, OpenAI Codex) to interpret instructions and handle tasks requiring judgment — triaging issues, fixing CI failures, updating docs.

How It Works

✍️
Write
Markdown + YAML frontmatter
⚙️
Compile
gh aw compile
📤
Push
.md + .lock.yml
🚀
Run
GitHub Actions
👀
Review
Human-in-the-loop

Key Properties

PlatformRuns inside GitHub Actions as standard workflows
AuthoringNatural language in Markdown with YAML frontmatter
SecurityRead-only by default; writes only through “safe outputs”
AgentsCopilot CLI (default), Claude Code, OpenAI Codex
LicenseMIT — fully open source
LanguageCLI written in Go (70.8%), JavaScript (27.6%)
Homepagegh.io/gh-aw

History & Origins

Created at GitHub Next by Don Syme and Peli de Halleux — from private R&D to 4,000+ stars in under a year.

Timeline

August 12, 2025
Repository created — private development begins at GitHub Next
January 12–13, 2026
”Peli’s Agent Factory” blog series launches — 19-part series documenting 100+ workflows
February 5, 2026
”Continuous AI in Practice” blog post published on the GitHub Blog
February 13, 2026
Official Technical Preview — announced on GitHub Blog and Changelog
March 25, 2026
v0.64.0 released — 313 total releases, 40 contributors, 4,167 stars

Creators

Don Syme

Principal Researcher at Microsoft Research, creator of F#, now at GitHub Next. Lead author of the official blog posts and the 19-part “Meet the Workflows” blog series.

Peli de Halleux

Researcher at GitHub Next. Creator of “Peli’s Agent Factory” — a collection of 100+ automated agentic workflows run continuously in the gh-aw repository and across GitHub internally.

Collaboration

A collaboration between GitHub, Microsoft Research, and Azure Core Upstream. Top contributors include Copilot (AI), dsyme, pelikhan, and 40 others.

”Waking Up to a Healthier Repository”

The founding vision: what does repository automation with strong guardrails look like in the era of AI coding agents?

”Imagine visiting your repository in the morning and feeling calm because you see:”
— Don Syme & Peli de Halleux, GitHub Blog, February 13, 2026

🏷️ Issues Triaged & Labelled

New issues automatically analyzed, categorized, and assigned to the right teams — semantically, not just by keywords.

🔧 CI Failures Investigated

Failed builds analyzed and correlated with recent changes, with proposed fixes ready for your review.

📝 Documentation Updated

Docs automatically aligned with recent code changes — no more stale READMEs or outdated API references.

🧪 Testing PRs Await Review

New pull requests that improve test coverage are ready for human review — small, focused, and incremental.

”All of it visible, inspectable, and operating within the boundaries you’ve defined.”
”A useful mental model: if repetitive work in a repository can be described in words, it might be a good fit for an agentic workflow.”

🏭 Peli’s Agent Factory — The Living Proof

A collection of 100+ automated agentic workflows running continuously in the gh-aw repository. These specialized agents handle documentation, style, cleanup, security, triage, and culture.

”These agents never take a day off, quietly working to make our codebase better.”

Continuous AI

The systematic, automated application of AI to software collaboration — augmenting CI/CD with judgment-based, context-dependent automation.

”GitHub Agentic Workflows hosts coding agents in GitHub Actions, to perform complex, multi-step tasks automatically. This enables Continuous AI — systematic, automated application of AI to software collaboration.”
Continuous AI = natural-language rules + agentic reasoning, executed continuously inside your repository
— GitHub Blog

Not a Replacement for CI/CD

”GitHub Agentic Workflows and Continuous AI are designed to augment existing CI/CD rather than replace it. They do not replace build, test, or release pipelines, and their use cases largely do not overlap with deterministic CI/CD workflows.”

The Gap Continuous AI Fills

CI/CD (Deterministic)Continuous AI (Judgment-Based)
Tests pass or failDoes documentation match implementation?
Builds succeed or don’tIs this issue properly triaged?
Linter flags violationsAre there performance anti-patterns?
Rules-based automationIntent-based automation
YAML configurationNatural language instructions
”CI isn’t failing. It’s doing exactly what it was designed to do. But many of the hardest and most time-consuming parts of engineering are judgment-heavy and context-dependent.”

The 30+ Year Vision

”The first era of AI for code was about code generation. The second era involves cognition and tackling the cognitively heavy chores off of developers.”
— Idan Gazit, head of GitHub Next
”This is the first harbinger of the new phase of AI. We’re moving from generation to reasoning.”
— Idan Gazit, head of GitHub Next

Humans Stay in the Loop

”Agentic workflows create an agent-only, sub-loop that’s able to be autonomous because agents are acting under defined terms. But it’s important that humans stay in the broader loop of forward progress in the repository, through reports, issues, and pull requests. With GitHub Agentic Workflows, pull requests are never merged automatically, and humans must always review and approve.”

The Six Pillars of Continuous AI

Six categories of repository automation that showcase what’s possible when AI handles judgment-heavy, context-dependent tasks — things that “would be difficult or impossible to accomplish with traditional YAML workflows alone.”

🏷️

Pillar 1: Continuous Triage

Automatically summarize, label, and route new issues. The agent reads incoming issues, understands their content semantically, applies appropriate labels, assigns to the right team members, and adds helpful summaries or requests for clarification.

Example workflows: Issue Clarifier, Auto-labeler, Priority Router

📖

Pillar 2: Continuous Documentation

Keep READMEs and documentation aligned with code changes. The agent reads function docstrings, compares them to implementations, detects mismatches, and opens PRs to update either the code or documentation.

57 merged PRs — 96% merge rate Glossary: 100% merge rate
🔄

Pillar 3: Continuous Code Simplification

Repeatedly identify code improvements and open pull requests. Pre-compiling regex patterns, removing dead code, simplifying complex conditionals, and extracting reusable functions.

88 merged / 103 proposed — 85% rate
🧪

Pillar 4: Continuous Test Improvement

Assess test coverage and add high-value tests. Small PRs daily, developers review incrementally.

~5% → near 100% coverage 1,400+ tests in 45 days ~$80 in tokens
🛡️

Pillar 5: Continuous Quality Hygiene

Proactively investigate CI failures and propose targeted fixes. Analyzes failed CI runs, correlates with recent changes, proposes fixes for flaky tests, identifies dependency conflicts, and suggests configuration improvements.

📊

Pillar 6: Continuous Reporting

Create regular reports on repository health, activity, and trends. Pulls from issues, PRs, commits, and CI results to synthesize cross-source insights.

”The value isn’t the report itself. It’s the synthesis across multiple data sources that would otherwise require manual analysis.”

📊 Real Metrics from Peli’s Agent Factory — Documentation Pillar

WorkflowMerged PRsProposed PRsMerge Rate
Daily Documentation Updater575996%
Glossary Maintainer1010100%
Documentation Unbloat8810385%
Documentation Noob Tester92143%
Slide Deck Maintainer2540%
Multi-device Docs Tester22100%
Blog Auditor5 passed, 1 flagged6 audits

Markdown-First Authoring

A fundamental shift from “programming your automation” to “describing your automation” — write your intent in Markdown, compile it to a hardened workflow.

”The concept behind GitHub Agentic Workflows is straightforward: you describe the outcomes you want in plain Markdown, add this as an automated workflow to your repository, and it executes using a coding agent in GitHub Actions.”

Workflow File Structure

📋 YAML Frontmatter

Configuration between --- markers:

  • Triggers (on:) — schedule, events, manual
  • Permissions — read-only by default
  • Tools — github, browser, web-search
  • Safe Outputs — pre-approved write operations
  • Engine configuration — copilot, claude, codex

📝 Markdown Instructions

Natural language task description:

  • What to analyze or investigate
  • Style and format guidance
  • Process steps (guidance, not rigid procedures)
  • Scope and specificity controls

The Compilation Flow

📄
workflow.md
Human-readable intent
⚙️
gh aw compile
Validates, hardens, restricts
🔒
workflow.lock.yml
Machine-executable
🚀
GitHub Actions
Runs as standard workflow

Both the .md source file and the .lock.yml compiled file are committed to the repository. The .md is the editable source of truth, while the .lock.yml is the compiled, security-hardened version that actually executes.

”Your workflows can range from very general (‘Improve the software’) to very specific (‘Check that all technical documentation and error messages for this educational software are written in a style suitable for an audience of age 10 or above’).”

Agentic vs. Traditional Workflows

A side-by-side comparison across 10 dimensions showing how agentic workflows differ from — and complement — traditional GitHub Actions.

Dimension⚙️ Traditional GitHub Actions⚡ GitHub Agentic Workflows
AuthoringYAML with explicit step definitionsMarkdown with natural language intent
LogicDeterministic if/then rulesAI-driven contextual reasoning
BehaviorFixed — same input = same outputAdaptive — interprets context flexibly
Decision MakingPre-programmed stepsAgent makes autonomous decisions
OutputsDirect writes with granted permissionsControlled “safe outputs” with validation
Security ModelPermissions granted to workflowRead-only by default, explicit safe outputs
Use CasesBuilds, tests, deployments, lintingTriage, documentation, code quality, reporting
ExecutionSteps run sequentially as codedAgent reasons through instructions
Error HandlingExplicit error handlersAgent adapts based on context
ReviewAutomated outputs applied directlyHuman review required for all mutations
”Traditional workflows execute pre-programmed steps with fixed if/then logic. They do exactly what you tell them, every time, in the same way. Agentic workflows use AI to understand context, make decisions, and generate content by interpreting natural language instructions flexibly.”
”Don’t use agentic workflows as a replacement for GitHub Actions YAML workflows for CI/CD. This approach extends continuous automation to more subjective, repetitive tasks that traditional CI/CD struggle to express.”

Supported Coding Agents

The same workflow format works across all supported engines — choose the coding agent that best fits your needs.

🤖

GitHub Copilot CLI

Default Engine

GitHub’s native coding agent. Each run typically incurs two premium requests: one for agentic work, one for guardrail checks. Models can be configured to manage costs.

engine: copilot
🧠

Claude Code

Supported

Anthropic’s Claude agent. Requires an ANTHROPIC_API_KEY secret added to the repository.

engine: claude

OpenAI Codex

Supported

OpenAI’s coding agent. Requires an OPENAI_API_KEY secret added to the repository.

engine: codex

Tools & Model Context Protocol (MCP)

All engines interact with repository data through the Model Context Protocol (MCP) — a standardized protocol for connecting AI agents to external tools and services.

yaml — Tool Configuration
tools:
github:              # GitHub MCP server
browser:             # Playwright-based browser automation
web-search:          # Web search capabilities

Repositories

Code, files, branches, commits

Issues & PRs

Create, read, comment, label

GitHub Actions

Workflow runs, jobs, logs

Key Terminology

Essential terms and concepts for working with GitHub Agentic Workflows.

Core Terms

Agentic WorkflowAn AI-powered automation written in Markdown that uses a coding agent to understand context, make decisions, and take actions within controlled boundaries.
Coding Agent / EngineThe AI model/provider (Copilot CLI, Claude Code, or Codex) that interprets natural language instructions and executes them using configured tools and permissions.
Continuous AI (CAI)The systematic, automated application of AI to software collaboration — natural-language rules + agentic reasoning, executed continuously inside your repository. Augments CI/CD, does not replace it.
FrontmatterThe YAML configuration section between --- markers at the top of a workflow Markdown file. Defines triggers, permissions, tools, engine, and safe outputs.
Safe OutputsPre-approved, validated GitHub operations that an agent is allowed to perform (e.g., create-issue, create-pull-request, add-comment). Agents run read-only by default; writes only happen through safe outputs.
Lock File (.lock.yml)The compiled output — a security-hardened GitHub Actions workflow file generated from the Markdown source via gh aw compile. Both files are committed to the repository.
MCPModel Context Protocol — a standardized protocol for connecting AI agents to external tools and services. Used for GitHub operations, external APIs, file operations, and custom integrations.
Threat DetectionAn automated security layer that analyzes agent outputs for prompt injection attempts, secret leaks, and malicious code patches. Runs after the agent job but before safe outputs are applied.
Agent Workflow FirewallA containerized, network-layer firewall built for AI agents. Restricts agent network access using domain allowlists, preventing agents from contacting unauthorized domains.
Protected FilesA static, rule-based protection layer guarding against supply chain attacks by preventing agents from modifying CI/CD configuration, dependency manifests, and agent instruction files.
Peli’s Agent FactoryA collection of 100+ automated agentic workflows running continuously in the gh-aw repository, created by Peli de Halleux as a practical demonstration of Continuous AI.

Design Patterns

ChatOps

Triggered by comment commands (/fix, /review)

DailyOps

Scheduled daily workflows (reports, cleanup)

IssueOps

Triggered by issue events

DataOps

Workflows that process and analyze data

ProjectOps

Project management automation

MultiRepoOps

Workflows spanning multiple repositories

Complete Workflow Examples

Real workflow files from official sources — copy, adapt, and run them in your own repository.

📊 Example 1: Daily Repository Status Report

The canonical example from the official GitHub Blog — generates a daily maintainer report as a GitHub issue.

markdown — daily-status.mdOfficial Example
---
on:
schedule: daily
permissions:
contents: read
issues: read
pull-requests: read
safe-outputs:
create-issue:
  title-prefix: "[repo status] "
  labels: [report]
tools:
github:
---

# Daily Repo Status Report

Create a daily status report for maintainers.

Include
- Recent repository activity (issues, PRs, discussions, releases, code changes)
- Progress tracking, goal reminders and highlights
- Project status and recommendations
- Actionable next steps for maintainers

Keep it concise and link to the relevant issues/PRs.

🔍 Example 2: Issue Clarifier

Automatically asks for additional details when new issues are unclear.

markdown — issue-clarifier.mdIssueOps Pattern
---
on:
issues:
  types: [opened]
permissions: read-all
safe-outputs:
add-comment:
add-labels:
---

# Issue Clarifier

Analyze the current issue and ask for additional details if the issue is unclear.

- Read the issue's title and body.
- If the report is missing key details (reproduction steps, environment,
expected/actual results), politely comment and request clarification.
- Add label `needs clarification` if applicable.
- If the issue is well-formed, add a `triaged` label and a brief
acknowledgment comment.

📖 Example 3: Documentation Updater

From Peli’s Agent Factory — achieved a 96% merge rate across 59 proposed PRs.

markdown — docs-updater.mdDailyOps Pattern
---
description: |
Reviews and updates documentation to ensure accuracy and completeness.
Compares docstrings with implementations and proposes fixes.
on:
schedule:
  interval: daily
push:
  branches: [main]
  paths: ['src/**', 'docs/**']
permissions:
contents: read
pull-requests: read
safe-outputs:
create-pull-request:
  title-prefix: "[docs] "
  labels: [documentation, automated]
  protected-files: fallback-to-issue
tools:
github:
---

# Daily Documentation Updater

Review the repository documentation and ensure it accurately reflects
the current codebase.

## Tasks

1. Compare function docstrings with their implementations
2. Check that README sections match actual behavior
3. Verify code examples still work
4. Identify outdated references or deprecated features
5. Propose specific, targeted fixes via pull request

## Guidelines

- Make small, focused changes (one topic per PR)
- Preserve the original author's voice and style
- Only fix genuine inaccuracies, not stylistic preferences
- Include a clear description of what changed and why

📰 Example 4: Continuous AI News Report

A minimal example from the Continuous AI in Practice blog — shows how simple a workflow can be.

markdown — news-report.mdMinimal Example
---
on: daily
permissions: read
safe-outputs:
create-issue:
  title-prefix: "[news] "
---

Analyze the recent activity in the repository and:
- create an upbeat daily status report about the activity
- provide an agentic task description to improve the project
based on the activity.
Create an issue with the report.

Key Quotes from Official Sources

Words from the creators and leaders behind GitHub Agentic Workflows.

”Custom agents for offline tasks, that’s what Continuous AI is. Anything you couldn’t outsource before, you now can.”
— Idan Gazit, head of GitHub Next
”Any task that requires judgment goes beyond heuristics. Any time something can’t be expressed as a rule or a flow chart is a place where AI becomes incredibly helpful.”
— Idan Gazit, head of GitHub Next
”In the future, it’s not about agents running in your repositories. It’s about being able to presume you can cheaply define agents for anything you want off your plate permanently.”
— Idan Gazit, head of GitHub Next
”Designing for safety and control is non-negotiable. GitHub Agentic Workflows implements a defense-in-depth security architecture that protects against unintended behaviors and prompt-injection attacks.”
”The PR is the existing noun where developers expect to review work. It’s the checkpoint everyone rallies around.”
— Idan Gazit, head of GitHub Next
”Treat the workflow Markdown as code. Review changes, keep it small, and evolve it intentionally.”
”Documentation is where we challenged conventional wisdom. Can AI agents write good documentation?”
”AI-generated docs need human/agent review, but they’re dramatically better than no docs (which is often the alternative).”
”Test coverage went from ~5% to near 100%. 1,400+ tests were written across 45 days for about ~$80 worth of tokens.”