Developers often face a dilemma: when working on complex projects, they must juggle multiple tasks—like adding features, reviewing code, or running tests—each requiring different AI assistance. Traditionally, these tasks had to be handled sequentially, leading to delays and cognitive overhead. In a recent update, GitHub Copilot introduces the ability to run multiple AI agents in parallel, allowing developers to manage several tasks at once without interference.
What Happened: A New Capability for Parallel Agent Sessions
As reported on the GitHub Blog, the GitHub Copilot app now supports running multiple agent sessions simultaneously. This feature allows users to initiate several AI-driven coding tasks—such as feature development, accessibility reviews, or automated testing—within the same project, each operating independently.
The key innovation lies in how these sessions are isolated. Each agent runs in its own Git worktree, a dedicated environment that ensures no interference between tasks. This means that one agent’s progress does not affect another, and users can switch between sessions seamlessly without losing context or having to restart tasks.
Key Facts and How It Works
At the heart of this functionality are three core concepts:
- Agent sessions: A session represents a single, end-to-end task you assign to Copilot, such as writing a function or generating test cases. Each session has a title and a progress indicator visible in the app’s sessions view.
- Git worktrees: Each session operates within its own isolated Git worktree. This isolation prevents conflicts and ensures that changes made in one session do not impact others.
- Context preservation: When a session is paused or switched, Copilot retains the full context of what was being worked on. This means developers can return to a session and resume exactly where they left off, without re-explaining their goals.
For example, in a project called tailspin-toys, a developer might initiate three distinct tasks:
- Build a funded sort feature.
- Conduct an accessibility review.
- Run automated tests on the codebase.
Each task begins independently, runs in its own environment, and progresses without blocking the others. The developer can monitor all sessions through the app’s sessions view, or even step away to grab coffee—knowing that the work continues uninterrupted.
Why It Matters: Productivity and Developer Experience
Before this feature, developers often had to wait for one AI task to complete before starting the next. This sequential workflow created bottlenecks and reduced overall efficiency. With parallel agent sessions, developers can now multitask more effectively, reducing the mental load of switching between tasks.
According to the original source, this capability transforms the developer experience by minimizing context switching and eliminating the need to constantly monitor AI agents. It shifts the focus from managing tools to actually doing work—something many developers have long sought.
Moreover, this feature aligns with broader trends in AI development tools. As highlighted in a previous Siriusity article, developers increasingly prefer tangible, visual interfaces over chat-based interactions. Parallel agent sessions in Copilot reflect this shift—offering a more structured, workflow-oriented experience.
These improvements are especially valuable in large-scale projects where multiple aspects of code quality must be addressed simultaneously. For instance, a feature might need to be implemented, tested, and audited for accessibility—all at once—without requiring a developer to switch between tools or restate their intentions.
Limitations and Open Questions
While the feature is a significant step forward, it is not without limitations. The original source does not detail how the system handles conflicts between tasks—such as when two agents generate conflicting code or when one task depends on another’s output. Since the sessions are isolated, there is no built-in mechanism for coordination or dependency management.
Additionally, the feature is currently best suited for small, independent tasks. Complex workflows involving interdependent steps or long-running processes may still require manual oversight or integration with other tools. The absence of a workflow orchestration layer means that developers must still manage task sequencing manually.
Another open question is how this feature interacts with existing codebases. While the isolation of worktrees prevents interference, it does not guarantee that generated code will be consistent with project-wide standards or architectural patterns. This could lead to inconsistencies if not properly reviewed.
What to Watch Next
As GitHub continues to evolve Copilot, several developments are likely to follow:
- Integration with GitHub’s AI-powered fuzzing tools: The AI-powered fuzzing tool already supports C/C++ projects. Future versions may allow parallel agent sessions to include security testing alongside feature development.
- Workflow automation with canvases: As discussed in a prior article, GitHub is introducing canvases—visual surfaces that allow developers to design and update workflows directly. These could soon support parallel agent execution as a built-in component.
- Scalability in large repositories: The ability to run multiple agents in parallel will be tested in larger codebases. How Copilot manages memory, performance, and session stability in million-line repositories remains to be seen, especially as highlighted in a previous case study.
These developments suggest that GitHub is moving toward a more integrated, workflow-centric AI experience—one where AI agents don’t just assist, but actively manage the flow of development work.
For developers, this means a more intuitive, less intimidating AI environment. The moment when AI stops feeling like a chore and starts feeling like a powerful, reliable partner is now within reach.
Sources & further reading
Featured image: Inside CSIRO Marine Laboratories Library, Hobart, Tasmania, 2005. Photo credit: Tony Rees by Tony 1212, CC BY 4.0, via Wikimedia Commons. Image source · License