Introducing GitHub Agentic Workflows: AI that runs your repo

[Music]

[MUSIC FADES INTO SWEEPING SOUND]

PELI DE HALLEUX: Hello everyone. My name is Peli de Halleux and I work for Microsoft Research, and today I’m going to talk about GitHub agentic workflows.

This is a project with GitHub, GitHub Next, and Azure Core, and it aims at using agents to automate all the things. By all the things, we mean automate the entire spectrum of the software development lifecycle, from testing, generating code, documentation, and much more.

We think that automation will be the true accelerator of agentic transformation, going from single developer productivity to 100x, 1000x multiplication. But we also think that this multiplication will come with order and orchestration and processes, and not with swarms and chaos. And this is what we plan to build.

In GitHub agentic workflows, we build agentic human processes powered by GitHub. We have automation through GitHub Actions. We build safety through a sandbox. And we have reasoning through GitHub Copilot CLI. We put the three together, and you have GitHub agentic workflows.

For example, let’s take a look at this process. It’s a research plan assign. It is a process that involves multiple agents and multiple human intervention.

It starts with deep research on a schedule that tries to solve a problem, like finding duplicate code. It generates a report. The report is inspected by a developer. The developer decides that the findings are pretty good, and it spawns another agent to turn that report into work items.

Now, an automated agent, IssueMonster, assigns them to Copilot so that they get turned into code. IssueMonster looks at any issue that’s tagged with cookie—it loves to eat cookies.

Then Copilot does the usual transformation to code with a dance with the developer, where the developer can add comments and reviews. Eventually, this becomes a pull request and gets merged.

This is an example of an agentic human process, where multiple agents and multiple humans were involved.
Now, agents are dangerous if they are left unattended. Whenever you do an agentic process, you may have adversarial strings entering your process through pull requests, issues, or web queries. Anything coming from an untrusted source may try to take over your agents.

We built a sandbox that deals with that. We built a lot of layers of security, and this is one of the aspects that’s very important at Microsoft Research. We have the time, the energy, and the people who are able to drill into these problems. This is a very important aspect of agentic workflows: safety.

To support the transformation to agents, we also changed the language. We converted the GitHub Action YAML file format to markdown, which is very popular for agents. In our file format, the front matter at the top is the old actions plus some agentic stuff, and then the second part is a prompt.

But of course, you don’t edit these files yourself—you use an agent. And you can see on the right that we are using a phone. The future of interacting with agents is going to be natural language-based, and you will not need a laptop. It could be a phone, it could be through voice. And this is also something that we’re exploring in GitHub agentic workflows.

Agentic human processes will not only change software development, they will change any process that uses information and reasoning. And by that, we mean non-developers such as people in marketing, sales, or operations.
Anybody who’s doing reasoning and using a Copilot today may want to automate this Copilot.
This project is a collaboration between Microsoft Research, GitHub Next, GitHub, and Azure Core. The project is open source, and you can find the link below.
Thank you for watching.