GitHub Copilot agents can now fix review comments and merge PRs

What happened

GitHub Copilot is an AI assistant that helps developers write and change code. Visual Studio Code, or VS Code, is a widely used free code editor from Microsoft. GitHub's changelog post of 1 October 2026 covers five VS Code releases shipped during September. GitHub says the goal was to support agent-driven work from writing the code all the way to merging it into the main project.

The largest change is called agent merge. When it is switched on in an active session, the agent handles reviewer feedback, failed checks, merge conflicts and reruns of automated workflows. This feature is in preview. A second preview feature, automations, lets a developer schedule tasks to run every hour, every day or every week, or start them on demand, from a template or from their own instruction.

Developers can now open a pull request directly from an agent session in the Agents window. GitHub says this works for sessions with Copilot, Claude or Codex agents. The developer can review and edit the title and description and choose draft and merge options before creating it. Agents can also run inside Dev Containers, which are prepared software environments, on remote machines reached over SSH, tunnels or the Windows Subsystem for Linux (WSL).

A feature called HydraFusion, in research preview, chooses the AI model and the way of working for each coding task automatically. GitHub says it is available to eligible users who switch on preview features. Other changes help developers keep many agent sessions organised, attach a GitHub issue or pull request to a chat, and continue a Codex conversation from the ChatGPT app inside VS Code.

The engineering behind it

Most software teams work through pull requests. A developer makes changes on a separate branch, then asks for them to be merged into the main code. Teammates review the changes and leave comments. Automated tests and checks, called continuous integration, run on every change. If two people edited the same lines, Git reports a merge conflict that someone must resolve by hand. Only when reviews pass and checks are green is the change merged.

Each of these steps takes time, and agent merge tries to automate the repetitive part of the loop. The agent reads a reviewer's comment, changes the code, pushes the change and waits for checks. If a test fails, it looks at the error and tries again. This only works well if the tests are good. An agent can make a failing test pass in the wrong way, so a human still needs to judge whether a fix is correct.

Giving an agent more freedom also raises a security question. An agent that runs commands on your computer could read files, use the network or use stored passwords. In a weekly update on 25 September, GitHub added local sandboxing in public preview, which limits what an agent can reach. On 7 October GitHub made it generally available. It says the sandbox restricts files, internet access and Git credentials, using rules the developer or organisation sets.

New models

The same weekly update of 25 September added several new AI models to Copilot. According to GitHub, Claude Opus 5.5 and GPT-6 Sol became available on the Pro+, Max, Business and Enterprise plans. GPT-6 Luna and Grok 4.7 became available on the Pro plan as well as the higher plans. GitHub also added assisted approvals in the JetBrains editors, which approve low-risk tool calls automatically and ask the user before higher-risk actions.

The changelog does not include any measurements of how well agent merge or HydraFusion perform. The features are described as preview or research preview, which means GitHub may change them before a full release. Developers should treat them as tools to try, with their own checks in place, rather than as finished products.

What it means in Nepal

The sources contain no information about Nepal, so this section is about skills. These tools automate parts of a workflow that professional teams already use: branches, pull requests, code review and automated tests. A developer who does not understand that workflow cannot tell whether the agent did the right thing. A developer who does understand it can use the agent to save time on routine steps and spend more attention on design and review.

The practical skills are clear. Learn Git well, including branching and resolving conflicts by hand. Write tests for your own code, because an agent's fixes are only as reliable as the tests that check them. Write clear pull request descriptions that explain what changed and why. Before letting any agent run commands on your laptop, limit what it can reach, and never leave passwords or keys in project folders.

Student projects are a good place to practise. A team project on GitHub, with each member working on a branch and reviewing others' pull requests, builds the same habits that these tools are designed around. Adding a simple automated test run to the project shows how continuous integration catches mistakes before they reach the main branch.