GitHub shipped its August 2026 update to Copilot in Visual Studio this week, and the headline feature is a small UI addition with an outsized implication: a Low, Medium, High dial for how hard the underlying model thinks before it answers. Reasoning effort has existed as an API parameter for a while now, tucked into request payloads for anyone calling a model directly. Putting it on the surface of the editor, as a control a developer flips per task rather than a setting buried in a config file, is the part worth paying attention to.
The instinct behind it is simple. Not every prompt deserves the same amount of deliberation. Renaming a variable across a file and untangling a race condition in a distributed job queue are both “ask Copilot” moments, but they do not cost the same in tokens or time, and treating them identically wastes one of the two. Giving the developer the dial, rather than having the model guess, is a bet that humans still know their own problem better than a router does.
What actually shipped
Per-task thinking effort. Supported models now expose Low, Medium, and High settings directly in the Copilot prompt box. Low favors speed and lower token spend for straightforward edits; High spends more of the reasoning budget on things like architecture decisions or a bug that only reproduces under load. This is the same knob as reasoning_effort in modern chat completion APIs, just moved from a payload field to a widget you touch on every request.
A Git agent that reviews before you do. Copilot can now review uncommitted changes or a set of commits before a pull request exists at all. Findings show up inline in the editor and in a navigable list in Git Changes, and you can keep chatting with Copilot to understand or push back on any single suggestion. It works against both GitHub and Azure DevOps repositories, which matters for teams that have not fully consolidated onto one host.
Org-published custom agents. GitHub and enterprise owners can now publish a custom agent once and have it show up, with its description and source labeled, in every repository’s agent picker across the organization. No more copying an agent definition file from repo to repo and hoping it stays in sync.
Of the three, the thinking effort dial and the pre-PR Git agent are the pair that actually changes daily behavior, because they both insert a decision point into moments that used to be automatic: how much to think, and whether to look before committing.
Try it: effort versus payoff
About the demo
This is an illustrative model of the tradeoff Copilot’s effort dial describes, not a call to any real API. Move the slider and pick a task to see how the shape of the tradeoff shifts.Pick a task and move the slider.
Set the task to “diagnose a race condition” and drag the slider down to Low. The model in that scenario is not actually dumber, but it spends less of its budget checking its own work, which is exactly where a subtle concurrency bug hides. Now set the task back to a plain rename and leave the slider at High, and you are paying reasoning-token prices for a change a linter could catch.
Where this is heading
My take is that the thinking effort dial is the more durable idea of the three, because it is a UI pattern, not a Copilot-specific feature, and UI patterns spread. Once one major coding assistant puts reasoning depth in front of the user as a per-request choice instead of a fixed backend policy, the rest converge on it, the same way autocomplete, inline chat, and agent mode all became table stakes within a release cycle of each other. Expect JetBrains and VS Code proper to pick up an equivalent control soon.
The practical next step is less about the dial and more about discipline: pair the new Git agent review with whatever CI checks already gate your pull requests, rather than letting an inline “looks good” from Copilot substitute for one. A review that happens before the PR exists is genuinely useful for catching things early, but it is still the same model grading its own homework, and it deserves the same skepticism you would give any other reviewer that never sleeps.