Pichai: Google Behind on Agentic Coding

Sundar Pichai has admitted that Google is “a bit behind” in the emerging area of agentic coding, particularly when it comes to tool use, instruction following, and long-horizon software development tasks. His comments were made during an interview on the New York Times Hard Fork podcast, shortly after Google’s recent I/O developer conference, and offered a more candid reflection on the company’s position in the fast-moving AI landscape.

Pichai explained that while Google’s AI models remain highly competitive in several areas, including text generation, multimodal inputs, voice, audio processing, and general reasoning, there are still gaps when it comes to more advanced coding workflows. In particular, he pointed to agentic coding systems that can manage extended tasks across large and complex codebases, where models must maintain context and make multi-step decisions over time.

He contrasted this with areas where Google has traditionally been strong, such as generating single-shot outputs or building relatively simple web interfaces. However, he acknowledged that this strength does not fully translate into the more complex, ongoing development environments that modern AI coding tools are increasingly expected to handle.

A key issue, according to Pichai, is that Google has not historically had the same level of developer-facing products that generate large-scale, real-world coding interaction data. He suggested that this lack of exposure meant Google missed out on valuable feedback loops that competitors have benefited from. These systems, where developers interact directly with AI coding tools, help generate continuous data that can be used to improve model performance over time.

He pointed to the broader ecosystem around competing tools, noting that some rivals have been more tightly integrated into popular developer workflows. This has allowed them to collect richer behavioural data from users actively building and editing code, which in turn strengthens their models in practical coding scenarios.

Pichai referenced examples such as Anthropic’s integration with tools like Cursor, suggesting that Google had not previously established a similar “surface area” for capturing developer usage at scale. This, he implied, contributed to the gap the company is now working to close.

However, he stressed that this is beginning to change. At Google’s I/O conference, the company introduced Antigravity 2.0, a standalone desktop application designed specifically for agent-based coding workflows. This tool is part of a wider push to expand Google’s presence in developer environments and build more direct feedback loops between users and models.

According to Pichai, internal adoption of these tools within Google is growing extremely quickly. He said usage is effectively “doubling every week,” with engineers across the company increasingly incorporating the models into real development work. He suggested this rapid internal uptake is already helping improve performance, as the system benefits from constant real-world usage.

He also described the scale of internal engagement as unlike anything Google has seen before, particularly in terms of how quickly teams are integrating AI tools into their daily workflows. This, he said, is helping the company “hill climb” faster in terms of model improvement and iteration speed.

The interview also came just after the launch of Gemini 3.5 Flash, which has been rolled out as the default model for Google’s AI Mode globally. Pichai acknowledged that the launch initially came with some limitations that frustrated users, particularly around pricing structures, usage caps, and early performance consistency.

He explained that some of these restrictions were deliberately introduced to prevent system instability during rollout, but added that Google expects to relax these limits as capacity improves. He also accepted that early feedback included valid criticism, describing some of the constraints as “rightfully a source of frustration” for users.

On model quality, Pichai admitted that newer releases can sometimes introduce regressions in specific areas, even when overall performance improves. However, he said many of these issues are typically resolved quickly through post-training improvements and rapid iteration cycles.

Taken together, his comments presented a more cautious and self-aware perspective compared with Google’s recent public messaging at I/O, where announcements such as Gemini updates and Antigravity were framed in a more confident, forward-looking way. In contrast, this interview highlighted the competitive pressures still facing the company, particularly in the developer-focused segment of AI.

He described the current environment as highly dynamic, with rapid progress across the industry and no clear long-term leader in agentic coding systems. Despite the acknowledged gap, Pichai expressed confidence that Google is actively building the infrastructure needed to close it, particularly through increased developer engagement and internal usage of its own tools.

Looking ahead, he confirmed that Google continues to invest heavily in its Gemini models, including broader deployment of Gemini 3.5 Pro in the near future. While he stopped short of claiming the gap has been closed, his comments suggested that Google sees developer tooling and agentic coding as a key battleground for the next phase of AI competition.