Google co-founder Sergey Brin has reportedly urged DeepMind employees to focus on improving AI agents. This directive comes as the Alphabet-owned company works to advance its automated coding and research capabilities. In a recent internal memo, Brin emphasised the need for urgency in closing the gap in agent-based AI systems designed to handle multi-step tasks, as reported by the Information. His comments come as Google assembles a dedicated team of researchers and engineers to strengthen its AI coding models, with the broader aim of automating parts of its own development processes, the report added.In the memo, Brin wrote, “To win the final sprint, we must urgently bridge the gap in agentic execution and turn our models into primary developers” of code.The report cited two people familiar with the development, who said the latest push was partly driven by Anthropic’s recent AI model releases. Researchers at Google DeepMind view Anthropic’s coding tools as ahead of the code-writing capabilities of Google’s Gemini models, the people said.The end goal is AI takeoff, or AI that can improve itself, which has been a particular focus for Brin. He has told staffers that improving Google AI’s coding abilities is a step toward that eventual goal, the report noted. Both Brin and Google DeepMind CTO Koray Kavukcuoglu have been directly involved with the strike team, highlighting its importance to Google’s leadership, one of the people told The Information.An advanced coding agent combined with AI that can solve math problems and run experiments could automate the work of AI researchers and engineers at scale. OpenAI already has an internal tool to help its AI researchers work more efficiently. It can also generate code for experiments related to AI models.
How Anthropic has forced Google and OpenAI to focus more on coding
Coding has been a key focus this year across frontier AI labs, as OpenAI and Google work to catch up with Anthropic, which has been ahead in code generation abilities. Last month, OpenAI said it would shut down its Sora video generator to focus on coding and enterprise work.Google is now placing more emphasis on models that write code for internal use, shifting from its earlier focus on tools for external customers. This approach requires training models on Google’s own codebase, which differs from the public datasets typically used for general-purpose coding agents.Models trained on internal Google code cannot be released publicly. However, they could help in developing improved models that may later be made available.In a statement to the Information, a Google spokesperson said the company had seen “tremendous adoption” of internal coding tools. “Their use has been turbocharging our model and AI tooling development—we’re really focused here,” the spokesperson said.The strike team is led by Sebastian Borgeaud, a research engineer at Google DeepMind who previously led pretraining efforts, according to the people. The team is working on improving performance for long-term coding tasks, such as building new software. These tasks are complex and often require models to read multiple files and understand user intent, the people said.
