Google Built a DeepMind Strike Team After Sergey Brin Decided Anthropic’s Coding Agents Were Winning
Google’s Emergency Coding Team Says the Quiet Part Out Loud: Anthropic Is Ahead
Google does not usually admit weakness in public, even by implication. That is what makes Monday’s reports about a DeepMind “strike team” so interesting. According to The Information, echoed by The Verge, Sherwood, and Capital Brief, Google DeepMind has assembled a special team to fix Gemini’s coding weaknesses after deciding that Anthropic’s tools are better. That is not normal corporate housekeeping. That is a red flare.
The core detail came from a memo attributed to Google co-founder Sergey Brin. In that memo, Brin reportedly wrote that “every Gemini engineer must be forced to use internal agents for complex, multistep tasks.” Another line, quoted by Sherwood, was even blunter: “To win the final sprint, we must urgently bridge the gap in agentic execution and turn our models into primary developers.” Companies say a lot in polished keynotes. Internal memos tend to be more revealing.
Why this story matters now
The timing matters. The reports landed on April 20, 2026, and they point to a sharp shift inside Google. For the past year, the public AI race has often been framed around chatbots, search, and multimodal demos. This story cuts through that noise. The real contest may now be about which company can build coding agents that handle long, messy software work without falling apart halfway through a project.
Capital Brief reported that the new DeepMind team is focused on “long-term coding tasks,” the kind that require a model to read through files, keep track of intent, and make useful changes across a codebase. That is a harder problem than writing a neat function in a benchmark. It is also the problem that matters if a company wants AI to do real engineering work inside a giant software organization.
What stands out is the competitive target. This was not framed as a response to OpenAI. It was Anthropic. According to the reporting, DeepMind researchers viewed Anthropic’s coding tools as outperforming Gemini. That says a lot about where the center of gravity has moved. Anthropic was once treated as the careful lab with strong safety instincts and good language models. In 2026, it is forcing Google to reorganize around code.
Anthropic’s lead looks less accidental by the week
That reading matches other signals from the past few months. Anthropic has pushed Claude Code hard, and the company has not been shy about using its own tools internally. Capital Brief noted that Boris Cherny, head of Claude Code, said in January that “pretty much 100%” of Anthropic’s code is written using AI. That number sounds almost provocatively high, but it does two jobs at once: it markets the product and it tells rivals that Anthropic treats coding agents as infrastructure, not a side project.
Google has been moving in the same direction, just with less momentum. Anat Ashkenazi, Google’s CFO, said in February that around 50% of the company’s code was being written by coding agents, according to the same report. Fifty percent would have sounded astonishing a year ago. Here, it reads more like a sign that Google knows the direction of travel but does not yet control the pace.
The gap is not only about model quality. It is about workflow. Brin’s reported instruction that engineers “must be forced” to use internal agents is revealing because forced adoption is usually what management does when a tool is strategically important but not yet habitual. In plain terms, Google seems to believe that better agents will come from heavy internal use, more training data tied to real engineering tasks, and tighter loops between product teams and model researchers.
The people involved tell their own story
The names attached to the strike team matter. Capital Brief said Google CTO Koray Kavukcuoglu was directly involved. It also said the effort was led by Sebastian Borgeaud, a DeepMind research engineer who previously led pretraining work. Those are not minor appointments. When a company pulls in senior technical leadership for a special project, it usually means the normal roadmap is not moving fast enough.
There is another detail that deserves more attention. The Information, via Capital Brief, said Google is putting more emphasis on models trained on internal company code. Those models cannot simply be shipped to the public because the data is too sensitive. Still, the work can improve later public models. That is a familiar pattern in frontier AI now: the most useful systems may appear inside big companies before outsiders ever see the cleaned-up version.
This is bigger than a feature race
The easy reading is that Google wants a better answer to Claude Code. That is true, but too small. The larger issue is whether coding agents become the training ground for self-improving AI systems. The Verge noted that Brin sees catching Anthropic in coding as a step toward building AI that can improve itself. That line matters more than the memo drama.
If the strongest models can read repositories, write patches, run tests, understand failures, and adjust on the next pass, they stop looking like chat interfaces with autocomplete attached. They start looking like junior developers who never sleep, then like something more capable than that. The author suspects this is why Google’s response sounds unusually urgent. Search mattered. Chat mattered. But coding agents may decide who gets to build the next layer of AI fastest.
There is also a less flattering interpretation for Google. The company has world-class research talent, vast compute resources, and one of the deepest software ecosystems on the planet. If it still needs a strike team in April 2026 because Anthropic is ahead on coding agents, that suggests structure is getting in the way. Startups can focus. Giants often coordinate. In AI, coordination can be a tax.
What to watch next
The next few months should answer two concrete questions. First, does Google turn this internal sprint into a visible jump in Gemini’s coding performance, especially on long-horizon tasks across multiple files? Second, does Anthropic keep widening the gap by turning internal usage into faster product improvement?
Bottom line: the most revealing AI story in the last 24 hours was not a flashy launch. It was Google, through leaked reporting and an unusually sharp memo, effectively conceding that Anthropic has the stronger hand in coding agents right now. When a company as large as Google starts rearranging senior talent around a rival’s advantage, the market should pay attention. Quiet panic inside DeepMind is not proof that Anthropic wins the decade. It is proof that the race has entered a new phase, and code is where the pressure is showing first.
For tool-by-tool comparisons, see our AI coding listings and the comparisons section.