Google turns Android Studio into an open platform: 40 rival agents inside as the battle shifts to the 'plumbing'

By El Ecosistema Startup (with DiarioBitcoin and Digital Trends) · October 28, 2026.

At the Android Dev Summit, Google announced that Android Studio, its official development environment for mobile apps, will no longer be tied to a single built-in assistant. The feature is called Bring Your Own Agent (BYOA) and is already rolling out in preview in Android Studio Rabbit 2 Canary. It includes a marketplace with more than 40 agents, among them Claude Agent (Anthropic), Codex (OpenAI) and Google's own Google Antigravity.

According to the announcement, the agents don't just suggest code: they can access the project structure, the emulator and Logcat (the system and app log). They can run the app and read the output to check their own work. They connect through the Agent Client Protocol (ACP), an open standard for linking agents to editors, according to Digital Trends. That is why the list keeps growing without Google having to build a custom bridge for each one.

The package has more pieces. Android Bench 2.0 evaluates models and agents on multi-day engineering tasks, measuring functionality, time and cost. Change Review reviews what the agent has written and explains the changes step by step. Android CLI, the command-line tool, leaves its unstable phase with official plugins for Claude Code, Codex and Antigravity.

The performance figures come from Google's internal testing. Terminal previews of Jetpack Compose (Android's UI framework) reportedly sped up interface tasks by 77% and cut token usage, the text units that models bill for, by up to three times. In an A/B test with Claude Opus 5.5, fixing memory leaks took 46 minutes with the CLI versus 88 without it, across 45 user journeys. Digital Trends stresses that there was only one run per configuration, so it is a promising demo, not a verdict.

The second front is APIs that let apps take part in the agent ecosystem. AppFunctions API (beta) uses the Model Context Protocol (MCP) to expose app functions to local or system agents. Agentic API (alpha) runs Gemini Nano on the device. Agent-to-UI (alpha) lets agents present interactive controls instead of text. Connected Apps Platform (preview) brings services to Google Search and Gemini; Google cites more than 50 partner integrations, including Adobe, Peloton and Experian.

On the hardware side, Google and XREAL unveiled XREAL AURA, wired glasses for Android XR. They weigh under 95 grams, have a 70-degree field of view and offload computing to an external accessory with a Snapdragon Reality Elite chip. They start at $1,099, with shipping in mid-November, according to the cited sources.

Our take: the underlying move is strategic, not technical. Google is betting that the value lies not in having the best agent, but in being the place where everyone works. It's the pattern we've pointed out in other pieces: the battle is shifting to distribution, integration and standards, and the winner is whoever controls the plumbing, not just the smartest model. The fact that ACP and MCP are open standards reinforces this reading. Google gives up exclusivity over the assistant in exchange for remaining the environment all Android code passes through and, with AppFunctions, the gateway through which agents reach apps.

Who wins: developers, who can choose and combine agents without being locked into one vendor. Anthropic and OpenAI too, as they gain distribution in an environment they don't control. Who is exposed: proprietary assistants that relied on being the default, and Google itself if its own agents can't compete on quality. An open marketplace forces competition on merit, and that's healthy.

There are caveats. Several of these capabilities are still in alpha, beta or testing, and the productivity figures are internal and lack full conditions. For industry context, the token bill has become a real problem: the article cites TechCrunch, which reported that Uber had exhausted its 2026 AI coding budget by April. And a 2025 randomized trial with open-source maintainers found completion times 19% longer with AI, even though participants believed they had worked faster. Measuring cost per task and comparing cohorts is more reliable than trusting the feeling of speed.

In the short term, this means more automation of routine programming work and an uncomfortable transition for junior profiles. In the long term, once verifying and orchestrating agents becomes cheap and standard, building software will be accessible to far more people. It's a small piece of that abundance: less time on technical friction, more on deciding what's worth building.