At the Android Dev Summit, Google presented version Rabbit 2 of its Android Studio development environment. The most important innovation concerns less individual convenience features than the fundamental stance: the IDE – the integrated development environment in which code is written, tested and shipped – can in future also be operated with AI models that do not come from Google. Claude and GPT, among others, are named. At the same time, Google is expanding the agentic functions.
From assistant to agent
Agentic functions differ from classic code completion in that the model does not merely display suggestions but takes on multi-step tasks: reading files, making changes across multiple locations, checking results. The human remains in the role of the one who commissions and signs off, instead of typing every line themselves.
In practice, this shifts the focus of the work. The more independently a tool makes changes, the more important traceable diffs, clean commits and a test base that reliably makes regressions visible become. Anyone who brings agentic tools into projects is in fact investing first in review capability.
Why model openness is the real signal
Until now, AI support in Android Studio was closely tied to Google's own models. That third-party models are now also possible is remarkable, because manufacturers traditionally understand their development environments as a sales channel for their own platform. The opening suggests that being tied to a particular model is no longer a reliable unique selling point.
Above all, one reason for this is plausible: the quality differences between the leading models shift at short intervals. A team that is committed to one model today may be working with an inferior solution in a few months without being able to switch. Development environments that block this switch lose attractiveness.
What decision-makers can derive from this
The news directly concerns only Android development. However, the underlying trend is broader and relevant for all software projects in which AI tools are used.
- Model choice becomes configuration, not a fundamental decision. If the environment supports multiple providers, a switch can be made without breaking the toolchain.
- Comparability becomes possible. Only when several models run in the same environment can it be seriously assessed which delivers the better results for which task.
- Dependencies shift. In future, the tie will be less to the model than to the environment, the workflows and the internal conventions.
Open questions to clarify before use
Model openness does not automatically mean that every provider may be used in every company. Anyone connecting third-party models should clarify in advance which code components leave their own premises in the process, what contractual commitments exist regarding processing, and whether the use is compatible with internal guidelines and the requirements of client projects. For projects with sensitive data or strict data processing requirements, this is not a side issue but a prerequisite.
The cost side is equally unresolved in many teams. Third-party models are usually billed via separate accounts. Anyone testing several models in parallel needs transparency about who uses which quotas and for what.
A realistic start
Instead of making a company-wide decision, a limited trial setup is advisable. A manageable project, two to three clearly described task types and an honest assessment of the results provide more insight than any feature list.
- Define task types where support actually saves time – such as refactorings, test coverage or recurring adjustments.
- Run the same tasks with different models and document the rework.
- Adapt the review process so that machine-generated changes undergo the same review as hand-written code.
- Clarify data protection and contractual questions conclusively before use goes beyond the trial.
What remains important is a sober view: agentic tools shift effort, they do not eliminate it. The writing time saved partly reappears as review and correction time. A benefit arises where this exchange turns out positive on balance – and that can only be measured, not assumed.
Assessment for web and software projects
For ongoing projects, the development above all means that the question of the right AI model loses weight, while the question of robust processes gains weight. Those who have tests, reviews and version control under control can swap models as soon as better ones become available. Those who lack these foundations will not become faster through agentic tools, but merely harder to control.
Sources
- Android-Entwickler dürfen Fremdgehen: Google öffnet Studio für Claude, GPT usw. — heise developer News