Post

We're super jazzed to be launching GitHub Agentic Workflows into technical preview today! "Generative AI" kind of fooled us into thinking that AI is synchronous, but async AI can have so much more value for us as developers. 🧵 github.blog/ai-and-ml/au...

Heuristics can only take you so far in automation. There's no linter in existence that can tell you if your code and documentation have drifted out of sync. Agentic Workflows make it stupidly simple to express a task you want to delegate, forever — and have it do that for you, forever.

Core to the behavior of Agentic Workflows is the notion of safe outputs. The body of a workflow is markdown, but the frontmatter is where you explicitly define what the agent can read and write, and those are enforced outside of the agentic run.

You can check out a zillion examples in our docs site: gh.io/aw-tp-bsky It can be for easy things like "triage my issues" and it can be for hard things like "find ways to optimize my code". If it's a task you'd express in English to an engineer, it might be a good agentic workflow!

We're seeing teams at GitHub use it for building tooling and workflows for themselves. We're seeing developer-adjacent personas at GitHub using agentic workflows to help them move information around. Software is a team sport and this helps everyone on the team, not just the engineers.

And we're particularly excited about using this to help open-source maintainers. When we showed this to @frenck.social, the maintainer of Home Assistant, he wrote a workflow that reads the traces for bugs and figures out if they originate in the HA codebase or not. True noise elimination for OSS.