❓ Have you found your self orchestrating multiple coding agents across different harnesses, screens, surfaces? How many at most can you track and monitor simultaneously? 😫
❓ Did you ever give up on approving permission asks and started running all of them on YOLO mode ? Please try not to do that, at least not in production, agents can easily get out control and dangerous! 😱
Although I love all the magic and efficiency AI Coding Tools are bringing to my life, I'm a bit overwhelmed recently, trying to achieve tens maybe hundreds of tasks! Some really deserve my full attention and active steering to achieve the best possible outcome, while for most, the judgement of a model is more than enough.
In order to bring intelligent and controlled automation to my Github Repositories, I hired octopus students 🐙 🐙 🐙
GitHub Agentic Workflows(GH-AW).
What is GH-AW?
Disclaimer: GitHub Agentic Workflows a.k.a. GH-AW is an open-source GitHub CLI extension developed by GitHub. It is announced in February 2026 by GitHub Next and Microsoft Research, still in public preview but already widely adopted by Microsoft itself as well as many other organizations.
GH-AW enables running the coding agents you know and love such as GitHub Copilot, Claude Code, OpenAI Codex, or Google Gemini in GitHub Actions. Consider them as intelligent systems embedded directly into your CI/CD 💡 They all run within pre-defined guardrails, using safe outputs and sandboxed execution.
Some of the use cases are :
- Repository Maintenance with Issue Triages, PRs Reviews
- Continuous Documentation maintenance and consistency
- Metrics & Analytics for daily reports and workflow health monitoring
- Test improvements, CI failure diagnosis, quality checks
Hmm, what does it really address ?
👍 Do you want to gather daily activity data, analyze all the PRs and issues, check workflow results, and creates status report in form of issues ? Enable Daily Repo Status
👍 Do you want to automatically assign issues labeled cookie to the Copilot coding agent every 30 minutes? Enable Issue Monster.
Above are only two examples from a sample family of reusable pre-baked workflows maintained by Github Next, called agentics. There're plenty of interesting ones such as CI Doctor to Monitor CI workflows or Grumpy Reviewer for On-demand opinionated code review by a grumpy but thorough senior developer.
These may be useful, what else ?
👍 Do you have an automation in mind that you could describe entirely in natural language? Then you can create custom agentic workflows for that !
Whether it’s fixing a simple bug, creating and running tests on a schedule, fetching data from the web to keep your documentation up to date, or something entirely different.
How does it work ?
Agentic Workflow: From Prompt to Execution
gh aw compile
┌──────────────────────┐ ┌──────────────────────┐ ┌──────────────────────┐
│ 1. YOUR PROMPT │ │ 2. GH-AW │ │ 3. GH-AW COMPILER │
│ │ │ CUSTOMAGENT │ │ │
│ /agentic-workflows │ │ Creates workflow.md │ │ Validates source │
│ "Create a workflow" │──► using its skill │──► Injects runtime │
│ │ │ │ │ Generates YAML │
│ │ │ - schedule │ │ │
│ │ │ - permissions │ │ workflow.lock.yml │
│ │ │ - safe outputs │ │ │
│ │ │ - task brief │ │ │
└──────────────────────┘ └──────────────────────┘ └───────────┬──────────┘
│
┌───────────────────── commit + push ────────────────┘
▼
┌──────────────────────┐ ┌──────────────────────┐
│ 4. GITHUB ACTIONS │ │ 5. GITHUB REPO │
│ │ │ │
│ Launches agent │ │ Issues / PRs / code │
│ Executes task │ ───────► │ │
│ Posts safe outputs │ │ │
│ │ │ │
└──────────────────────┘ └──────────────────────┘
Why is this really a thing , can't I just build my own custom agents ?
Yes of course you can - sky is the limit ! But enforcing what they can do access and affect can can quite trickly. GH-AW runs within an isolated sandbox with Read-Only permissions, prevents Prompt Injection and is authorized to write only what you let.
When to use which: Github Actions VS Github Agentic Workflows
⚠️ A better question would be "How to use both together". 💡
Agentic workflows are complementary; your deterministic build, test, and release pipelines stay unchanged. Github Actions requires YAML configuration for fully deterministic systems. Agentic workflows use AI to interpret natural language instructions in markdown. Think of it as Continuous AI alongside CI/CD: a new automation layer in GitHub Actions for tasks where exact reproducibility doesn’t matter.
How can I start ?
A github repo to experiment some pre-built and 2 custom workflows
Disclaimer: Please don't hate me in case of errors 😃 and please feel free to submit issues or contribute 🤗Are you looking something for advanced, a zero to hero style workshop ? gh-aw-workshop
Last but not least, here comes the main source of all information and updates GitHub Agentic Workflows 🚀 🚀 🚀
Top comments (0)