GitHub Copilot is becoming part of more than the coding process. New integrations allow developers to start work from team conversations, delegate tasks to AI agents, and review the results. The opportunity is clear, but so is the need for good engineering judgment.
Imagine a bug report arriving in a Microsoft Teams thread. With the new @GitHub integration, a developer can use that conversation to investigate the issue, initiate a change, and open a pull request. The team still reviews, tests, and decides whether to merge the result.
GitHub’s recent announcements point in the same direction. Copilot is available in JetBrains Air for delegated coding tasks, while a public preview brings it into Xcode’s built-in coding assistant. The @GitHub integration in Slack and Teams is also generally available, making it possible to move from a conversation toward an issue, implementation, or pull request. Availability and setup requirements still differ between these tools.
For development teams, the important question is not how many new interfaces Copilot appears in. It is how these capabilities change the steps between identifying a problem, making a change, and deciding that the change is ready.
Many developers first met GitHub Copilot as a tool that suggested code while they typed. Agent-based workflows can go further. A developer describes an intended result, and an agent may inspect repository context, propose changes, and return a pull request for review, depending on its tools and permissions.
This changes what developers need to pay attention to. The quality of the task description, the context available to the agent, and the ability to understand its output all matter. Delegating a task does not remove the need to understand the problem or own the result.
What do you see as the biggest change for developers when AI agents can take on coding tasks rather than simply suggest code?
The biggest change is that we’re moving from writing code with AI assistance to delegating actual development tasks to AI agents.
That sounds great, and it is, but it also changes what we need from developers. Understanding architecture, security, maintainability and knowing how to evaluate the result becomes even more important.
An AI agent might write the code, but you’re still responsible for what ends up in production.
I think that’s where the real shift is. Less time spent typing code, more time spent making sure we’re building the right thing, in the right way.
An AI-generated change can look plausible while still missing a requirement or introducing a security issue. A passing test does not necessarily mean the change fits the architecture or will remain easy to maintain. Teams therefore need a clear understanding of who is responsible for reviewing, approving, and deploying agent-produced work.
GitHub has recently expanded the ways organizations can manage Copilot code review settings through APIs, including at enterprise and repository levels. These controls can help teams apply consistent policies, but they do not replace professional judgment about the code itself.
What should development teams pay particular attention to when using AI coding agents, especially when it comes to code quality, security, and human oversight?
I wouldn’t treat a pull request from an AI agent any differently from one written by a developer. The same quality and security standards should apply.
But AI agents can produce much more code, much faster, and realistically, not every line will receive the same level of human attention. That makes automated quality and security gates more important than ever.
Teams need proper CI/CD pipelines, using GitHub Actions, Azure DevOps or similar, with automated tests, static code analysis, dependency and vulnerability scanning, and secret detection. These checks should be mandatory before merging.
But a green pipeline doesn’t necessarily mean the code is good. Does it fit the architecture? Is it maintainable? Does it actually solve the problem? And don’t blindly trust tests generated by the same agent that wrote the code.
Ultimately, a developer who understands the system should review and approve the pull request.
You can delegate the work, but not the responsibility.
GitHub’s Copilot impact dashboard now includes engagement with Copilot Chat and the VS Code Agents window. This gives administrators a fuller picture of which capabilities developers use. But adoption metrics alone cannot show whether code is better, whether a team spends less time on rework, or whether delivery is more reliable.
Teams can start by comparing:
Remember to account for differences in task complexity when comparing the results.
Using AI coding agents effectively involves more than writing prompts. Developers need to work with repository context, judge generated changes, and understand what tools an agent is allowed to access. Technical leaders also need a common approach to responsibility, review, and learning across the team.
What skills should developers prioritize, and how would you recommend that a team starts experimenting with AI agents?
I’d start with something small and well-defined, like adding unit tests to an existing component that doesn’t have enough test coverage yet. It’s a task where the expected behaviour is already known, and you can easily review what the agent produces.
In my training sessions, I often emphasize that getting code generated is the easy part. Understanding whether it’s correct and actually useful is where the real engineering work comes in.
I’d measure one thing: how much time did we actually save? Not just the time it took the agent to generate the tests, but including the time developers spent reviewing, fixing and integrating them.
If the end result is good and the team genuinely saved time, you’ve found a useful application for AI agents. From there, you can gradually move towards more complex tasks.
The next step is not necessarily to automate every part of development. It is to understand where AI can contribute and make sure developers have the skills to guide, verify, and govern its work.
Glasspaper offers courses for teams at different stages of adopting AI-assisted development:
Want to work confidently with AI agents? Explore Glasspaper’s courses in GitHub Copilot and AI-assisted development, and build the skills to guide, verify, and govern the work agents produce.