GitHub and Google turn their developer tools into control rooms for swarms of AI agents

Two of the biggest players in developer tooling have reorganised their environments this week around AI agents that write code on their own. At its GitHub Universe conference, GitHub presented a revamped Copilot app that supports not only coding work but also specifications, design discussions and telemetry. InfoWorld describes how the new Copilot Projects feature can divide a broader goal spanning several repositories among agents working at the same time, with context from repositories, design mock-ups, Slack conversations and meeting notes. The result comes back as pull requests for review. A new Max tier for code review has several agents using models from different providers assess a change, after which an overarching agent merges their findings.

Analysts interviewed by InfoWorld see benefits for large modernisation jobs, such as Java or .NET upgrades across many repositories, but warn of conflicts when agents modify the same files. They point to a 2026 study in which agents succeeded roughly 30 percent less often when they had to collaborate. Review also risks becoming a bottleneck and costs can rise; companies would do well to set limits per project, they say.

Google took a similar step at the Android Dev Summit in Mountain View. Android Studio Rabbit 2 is now stable and includes Bring Your Own Agent: through the ACP protocol developers can choose from more than forty agents, including OpenAI's Codex, JetBrains' Junie and Google's own Antigravity, the Android developers blog reports. These get access to the emulator, the build system and Logcat. Heise lists further additions: a manager for Android skills, agents that track down memory leaks with LeakCanary, a /review command in the command line and a new version of Android Bench, which compares the performance and cost of models.