◀ Playbook index
NO.16

Agentic Workflow

Updated: 2026-10-26

In a nutshell

Agentic Workflow is a mechanism where AI agents automate work on top of GitHub Actions.

Triggered by cron, issue, or PR events, agents read context, make decisions, and take action.

How it works

Overview

  • AI agents automate work on top of GitHub Actions
  • Now in public preview (announced June 11, 2026)
  • No longer needs a PAT — runs with the built-in GITHUB_TOKEN

Components

  • ⚡ Trigger — same as Actions: schedule (cron) / issue / PR / comment / @mention, etc.
  • 📄 Description — write “what you want done” in natural-language Markdown in .github/workflows/NAME.md
  • 🤖 Agent — Copilot / Claude / Codex and others launched on a runner
  • 🔌 MCP / tools — agents call GitHub API and external services to do real work
  • 🛡️ safe-outputs — read-only by default. Writes go only through verified boundaries like PRs, comments, and labels

What can it do?

Use caseWhat it does
🏷️ Issue TriageReads new issues and automates labeling, priority assessment, and assignee suggestions. Requests clarification if ambiguous
📚 Wiki GeneratorPeriodically scans the codebase and auto-generates / updates Wiki / README
🔍 CI Failure AnalystAnalyzes failed CI runs, identifies root causes, and comments on the PR
📝 Docs MaintainerDetects stale links, examples, and API descriptions, then opens a fix PR aligned with the latest implementation

🧪 The sample collection githubnext/agentics has ready-to-run templates. Copy-paste to get started.
📝 Reading the developer blog Peli’s Agent Factory first gives you a good feel for how it all works.

Getting started (4 steps)

  1. Step 1 — Install the CLI extension — gh extension install github/gh-aw
  2. Step 2 — Write Markdown — Add trigger and instructions to .github/workflows/NAME.md
  3. Step 3 — Compile — Run gh aw compile to generate NAME.lock.yml
  4. Step 4 — Commit & push — Push including the lock file; Actions will trigger automatically from then on
---
on:
  issues:
    types: [opened]
permissions: read-all
safe-outputs:
  add-comment:
---

# Issue Clarifier
Read newly opened issues. If the requirements are ambiguous,
post a comment requesting clarification on these 3 points:
- What is happening / reproduction steps / expected behavior
Do nothing if the issue is already clear.

⚠️ Commit the lock file to Git. It is the actual running Actions workflow and the target of audit and review.

Slides 日本語