Copilot Automations now learn from outcomes and self-improve

Copilot Automations can now learn from their outcomes and propose improvements to themselves, helping them produce better results over time.

It’s easy to create an automation, but harder to build one you can trust to consistently produce high-quality results. Reviewing its output and improving its prompt by hand takes time. Self-improvement uses feedback from real outcomes on GitHub to make that process easier.

How it works

Your automations leave a trail of signals on GitHub that show how well they’re working. For example:

• An issue triage automation can learn from whether people change the labels it applied.
• A code review automation can learn from whether developers accept its comments.
• A performance optimization automation can learn from whether maintainers merge the pull requests it opens.

Self-improving automations use these signals to propose improvements to their prompts. You stay in control: each proposed change comes to you for review before it’s applied.

How Copilot creates a self-improvement loop

When you ask Copilot to set up a self-improving automation, it can create a group of automations that work together. For example, for a custom code reviewer, Copilot might create:

• A review automation that runs when a pull request is opened, reviews the changes, and posts comments.
• A weekly improvement automation that analyzes how developers responded to those comments and proposes refinements to the review automation.

When the improvement automation proposes a change to the original automation’s prompt, tools, or triggers, GitHub notifies you. You can review and accept or reject the proposed change before it is applied.

To learn more, see about Copilot Automations.

Copilot Automations is available with Copilot Pro, Pro+, Max, Business, and Enterprise plans.

Join the discussion within GitHub Community.