Alphabet is reorganising its artificial-intelligence empire. To catch Anthropic and OpenAI in coding, it may also need to buy the battlefield.

For a company that helped invent much of modern artificial intelligence, Google has spent an uncomfortable amount of the generative-AI era looking as though it were chasing companies it helped inspire.

That paradox has long been explained as a failure of execution rather than science. Google had transformers, world-class researchers, custom AI chips, enormous computing infrastructure and perhaps the richest distribution network in technology. What it lacked was the single-mindedness of a startup whose survival depended on shipping the next model or product before everyone else.

Alphabet’s latest leadership reshuffle is an attempt to change that.

Demis Hassabis, the co-founder of DeepMind, is stepping back from the day-to-day management of Google DeepMind to become its chairman and Alphabet’s chief scientist. Koray Kavukcuoglu, DeepMind’s chief technology officer, is assuming greater operational responsibility and will report directly to Sundar Pichai, Alphabet’s chief executive. At the same time Jeff Dean, one of the architects of Google’s modern computing and AI infrastructure, is leaving after 27 years, alongside several other senior researchers.

The symbolism matters. Mr Hassabis is one of the outstanding scientific leaders of his generation. But the question confronting Google today is increasingly not what intelligence is possible, but how quickly it can be turned into products people use every day.

Those are different jobs.

From laboratory to war room

DeepMind was built to pursue difficult scientific problems over long horizons. That culture produced extraordinary results, from AlphaGo to AlphaFold. Yet a research laboratory optimises for discoveries; a consumer-software company optimises for iteration.

The generative-AI market increasingly rewards the latter.

OpenAI and Anthropic can alter models, interfaces, system prompts, tool use and agent behaviour in response to customer feedback with relatively little organisational friction. Google, by contrast, must coordinate Gemini with Search, Android, Workspace, Cloud, YouTube, advertising and an assortment of developer products. A technically elegant decision for a general-purpose model may not be the right decision for a coding agent that developers expect to improve every few weeks.

Putting operational authority closer to Mr Pichai may therefore accelerate Gemini. It may also clarify an ambiguity that has dogged Google’s AI strategy: whether DeepMind is principally an advanced-research institution or the engine room of a commercial platform.

But management charts do not write software.

Google’s more serious challenge is that the contest in AI coding is moving beyond the model itself.

The model is becoming only one layer

For much of the generative-AI boom, competitive comparisons centred on benchmarks: whose model was best at mathematics, reasoning or programming?

That is becoming insufficient.

A coding model sits inside an increasingly elaborate system consisting of a terminal, file access, source control, testing, browsers, development environments, tools, permissions, memory and repeated interactions with the programmer.

In other words, the important product is becoming the coding agent, not merely the language model underneath it.

This distinction helps explain Anthropic’s success with Claude Code. Its advantage is not necessarily that Claude will permanently outperform Gemini or OpenAI’s models on every coding benchmark. Rather, Anthropic has created a coherent developer workflow around the model.

Evidence from open-source software already suggests that such agents are spreading unusually quickly. One recent study of more than 129,000 GitHub projects estimated coding-agent adoption at roughly 16-23% among the projects examined, despite the category being only months old. Another dataset identified nearly a million agent-authored pull requests across more than 100,000 repositories.

That matters because workflows produce habits, and habits produce lock-in.

A developer who merely calls an API can switch models easily. A developer who has built repositories, instructions, permissions, plugins, tools, testing procedures and team conventions around a particular agent is much harder to move.

The danger for Google is therefore not simply that Claude writes better code this quarter. It is that Claude Code—or OpenAI’s Codex, or another rival—becomes the default operating layer through which software is created.

Google has seen this movie before.

Search was extraordinarily valuable, but Google still concluded that it needed Chrome. Web services were extraordinarily valuable, but mobile computing still required Android. Controlling the underlying service did not eliminate the strategic importance of controlling the interface through which users reached it.

Coding agents may become a similar interface.

Google has already started shopping

Alphabet appears to recognise at least part of this problem.

In 2025 Google paid roughly $2.4bn for licences to Windsurf technology and hired several of the coding startup’s senior employees, without acquiring the company itself.

More recently Google has reportedly been discussing a transaction worth more than $1.5bn involving technology and employees from Mechanize, another AI-coding startup. The proposed structure would again rely on technology licensing and hiring rather than a conventional acquisition.

Such deals make sense. They may be easier to complete in an era of heightened antitrust scrutiny, and they allow Google to absorb scarce engineering talent without buying an entire corporate structure.

But they may also reveal the limits of Google’s approach.

Talent is valuable. Technology is valuable. Neither necessarily brings the most strategically important asset: the developer relationship.

A coding-agent company with millions of developers using its interface every week possesses something that cannot be recreated merely by hiring its best researchers. It owns a distribution channel, a stream of behavioural feedback and the beginnings of an ecosystem.

Google should therefore consider a more ambitious acquisition.

Buy the workflow, not just the engineers

An attractive target would resemble OpenCode, an open-source coding agent that works across terminals, IDEs and desktop environments and supports models from multiple providers. OpenCode is notable precisely because it is not tied to one model company. Users can combine the agent layer with models and services from different vendors.

The specific company matters less than the strategic properties Google should seek.

A useful acquisition target should already have meaningful developer adoption, own the agent interface rather than merely a model, support a growing ecosystem of tools and integrations, and be capable of remaining relatively model-agnostic.

That final point is crucial.

The worst thing Google could do after buying such a company would be to convert it into another Gemini-branded product, require Google accounts, optimise it exclusively for Gemini and subject every feature to a maze of internal product reviews.

That would destroy much of what it had bought.

Instead Google should treat the coding agent as an independent platform.

Gemini should be the preferred model only when it deserves to be. Claude, OpenAI models and open-source alternatives should remain available. Google Cloud should provide the compute, security, storage and enterprise infrastructure underneath them.

Such an arrangement would look strange by the standards of the previous software era but quite sensible in the AI one.

If Gemini becomes the best coding model, Google wins through the model.

If Claude remains better, developers might still run it through a Google-owned agent or on Google Cloud.

If open-source models become commoditised and competitive, Google can provide the infrastructure on which they run.

The goal would not be to predict which model wins. It would be to own more of the layer through which all of them are consumed.

An old platform strategy in a new disguise

This would fit Google better than it may initially appear.

Google has rarely required every layer of a technology stack to be proprietary. Android is open source. Chrome is built around Chromium. Google Cloud happily runs databases, operating systems and software created by others.

The company’s great strategic skill has often been to identify a layer that becomes widely used and then position Google services around it.

AI coding offers a similar opportunity.

At present Google’s developer-AI portfolio risks becoming too fragmented. Gemini Code Assist, Gemini CLI, Jules, AI Studio and other products attack overlapping parts of the developer workflow. Each may be individually useful, but abundance can become a liability when a rival offers one product with one name and one clear mental model.

Google does not need another coding feature. It needs an obvious place where developers go to build software with agents.

That platform should then become the integration point for Gemini, Google Cloud, source control, testing, deployment, security and third-party tools.

An acquisition could accelerate this process by years.

The antitrust problem

There is an obvious obstacle. Alphabet is already one of the world’s largest technology companies, and regulators in America and Europe would scrutinise any attempt to acquire an emerging developer platform.

That may explain Google’s preference for arrangements such as the Windsurf deal, in which technology is licensed and teams are hired without control of the target company.

Google may therefore need to be inventive.

It could acquire a minority stake while guaranteeing operational independence. It could sign a long-term commercial partnership. It could finance an open-source foundation. Or it could purchase a company while contractually preserving multi-model access and open protocols.

What matters is not formal ownership so much as strategic alignment.

Google needs a developer-agent ecosystem tightly connected to its infrastructure without suffocating the neutrality that makes such an ecosystem attractive.

Google does not need to win every benchmark

None of this means Gemini must become unquestionably superior to every rival.

Indeed, Google has structural advantages that OpenAI and Anthropic do not.

It has Search, Android, Chrome, Gmail, Workspace, YouTube and billions of existing users. It owns enormous data-centre infrastructure and its own TPU accelerators. It operates one of the world’s largest cloud platforms.

This gives Alphabet an unusual strategic position.

OpenAI and Anthropic generally need users to decide to come to them. Google can put AI in products people already use.

As long as Gemini remains within the frontier group of models, distribution can compensate for modest differences in capability. And if model performance eventually commoditises, Google can still make money providing the infrastructure on which competing models run.

But there is an important exception.

Google cannot afford to ignore a new computing interface merely because its existing businesses remain strong.

Microsoft understood this when it bought GitHub. Google understood it when it built Chrome and Android. The owner of an important interface can influence what sits above it and what sits below it.

Coding agents may become one of those interfaces.

The software engineer of the future may not spend most of the day typing code. Instead, he may instruct several agents, inspect their work, approve changes and manage increasingly autonomous software systems.

Whichever company owns that environment will occupy an unusually privileged position in the technology stack.

Google’s leadership reshuffle is therefore sensible but incomplete.

Moving Mr Hassabis towards long-term science and giving Mr Kavukcuoglu more operational authority may make Gemini faster and more commercially focused. But organisational efficiency alone will not close the gap with rivals that already own powerful developer workflows.

Alphabet should complement its management changes with a platform strategy.

It should acquire, invest in or tightly integrate with an independent coding-agent company of sufficient scale; preserve its multi-model architecture; make it a first-class citizen of Google Cloud; and resist the temptation to turn it into yet another Gemini sub-brand.

The strategic objective is not merely to build the smartest coding model.

It is to ensure that, five years from now, when a programmer begins work, the first AI environment he opens has Google somewhere in the stack.

For a company that once turned a search box into one of the most valuable positions in technology, that should be a familiar ambition.