GitHub’s adaptive review gives teams a new training exercise
GitHub put an Adaptive setting for Copilot code review into public preview on November 17. When a reviewer selects it, Copilot picks a review effort for the pull request and shows which level it used. Teams teaching developers to work with AI reviewers now have a concrete exercise: review a small change, read the effort label, then decide what a human still needs to check before approving it.
What changed in a pull request
According to the GitHub changelog, Adaptive chooses among Lite, Balanced and Max based on its assessment of the change's complexity and risk. The release also describes using context from related repositories. The two capabilities answer different questions. The effort label describes how the automated reviewer approached a change. A reference to another repository can help explain where it looked for context.
For a first workshop, take a sample pull request that changes the display of a delivery date on an order confirmation. On the surface it is a wording change. Elsewhere, a customer service template may repeat that wording. The trainer can supply both files as part of an exercise without using any customer records. A participant who knows the business process may notice the second use even if the automated review is quiet.
What to have a learner hand back
Keep the first deliverable small enough to read at the end of the session. Ask for the original requirement and one sentence recording the effort level shown for the review. Then ask the learner to cite the exact part of the proposed change that addresses the requirement. That last step matters. A summary such as “the change looks reasonable” gives the next reviewer nothing they can inspect.
Suppose the requirement says customers should see a dispatch date. The customer opens the order and sees Friday, the day the parcel should arrive. The warehouse means Wednesday, when it expects to hand the parcel to the carrier. Give the learner that example order. Ask them to underline the word in the requirement that settles which date belongs on screen and to describe a check that would catch the mix-up.
Run the same exercise with another reviewer
- Give two people the same sample change and the same written requirement. Ask them to note what they would verify before opening the generated review.
- Open the Copilot result. Record its effort label and pick a specific comment the group can evaluate from the available files. If there are no comments, record the fact and continue with the original checks.
- Compare the reviewers' final decisions. Where one person would approve and another would ask for a revision, find the sentence in the requirement that would settle the disagreement.
This approach lets a manager see where a course needs more time. Confusion about dispatch and delivery calls for a clearer business example. Trouble locating the evidence behind a generated comment calls for practice navigating a change. Running another model purely to produce more comments would do little for either learner.
