The tool base in software development is currently shifting every six months.According to Google DeepMind chief Koray Kavukcuoglu, the next model generation, Gemini 4, is set to appear significantly earlier than at the end of the year.According to him, the model is already in post-training – that is, in the phase in which a pre-trained language model is aligned to behaviour, safety and specific tasks through additional fine-tuning.
What is particularly notable is where the model is already in use: internally in the coding tool Antigravity.The next Gemini stage will therefore not first become visible as a chatbot demo, but in a development environment.That says a lot about which use case Google considers business-critical.
From the AGI mission to a product forge
Kavukcuoglu is visibly distancing himself from the debate that his predecessor Demis Hassabis shaped.He considers the question of general artificial intelligence – that is, a system that achieves human capabilities across arbitrary task areas – to be the wrong discussion.What matters more, he says, are agents that can be trusted.
That is more than a rhetorical shift.It describes a change of course: a research laboratory with a long-term mission has become a department that thinks in product cycles.This is flanked by two developments pointing in the same direction – the quiet discontinuation of Gemini 3.5 Pro and the departure of numerous leading researchers to OpenAI and Anthropic.
For companies that build AI functions into their processes, this shift can certainly be read as positive.Product orientation generally means more stable interfaces, clearer roadmaps and a stronger focus on measurable tasks instead of benchmark records.But it also means: models are replaced faster, and individual variants disappear without a big announcement.
Trustworthy agents are the real topic
The term agent describes an AI system that not only answers but independently executes work steps: reading files, changing code, triggering tests, checking results.This is exactly where it is decided whether AI saves effort in a project or creates additional review effort.
Trustworthiness is not a soft criterion here, but a technical one.It includes traceable changes, limited permissions, reproducible results and the ability to undo every step.An agent that works in a repository needs the same controls as a new employee during onboarding – only automated.
The fact that Google is testing the next model generation first in its own coding tool fits this logic.Development environments provide hard feedback: code compiles or it does not, tests pass or they do not.That is a significantly stricter review framework than a free conversation.
What this means for ongoing projects
Anyone already using AI assistants in development should soberly assess their own dependence on individual models.A few points that have proven themselves in practice:
- Plan for interchangeability: Route access to models through your own abstraction layer so that a change does not trigger rebuilds throughout the entire application.
- Pin model versions: In productive workflows, do not point to the latest variant in each case, but to a tested version – with a defined upgrade path.
- Factor in deprecations: The fact that a model variant is discontinued without much communication is a realistic scenario and belongs in the risk assessment.
- Make results measurable: Your own test cases from real project work say more about a model's suitability than general performance comparisons.
- Limit permissions: Agents should only be able to access the repositories, branches and systems they need for the specific task.
Review remains a human matter
The more powerful the assistance systems become, the more important the question of who bears responsibility for the delivered code becomes.An agent can create a pull request – the technical assessment of whether the change fits the domain model, the security architecture and the maintenance goals remains with the people on the team.
This pays off especially in TYPO3 and other CMS projects with long lifecycles.There, the most expensive errors do not arise in individual functions, but in decisions about data structures, extensions and update capability.AI tools accelerate implementation, but they do not replace architectural decisions.
Speed as a permanent state
The announcement of an earlier release is above all a signal about the market environment: the gap between model generations is getting shorter, and competition for developer teams is shifting from demonstrations to integrated tools.Anyone building a toolchain today should design it so that the next generational change is a configuration matter and not a project.
For web and software projects, this specifically means: the benefit of AI assistance does not arise from choosing the latest model in each case, but from clean interfaces, clear review processes and robust tests that safeguard every change.Teams that have these foundations can evaluate new models in days instead of weeks – and switch back again without risk.The strategic lever therefore lies less in the model itself than in the question of how controlled a company opens its development processes for automated contributions.