Add preferred source
Alphabet CEO Sundar Pichai said he now uses cloud-based AI agents to summarize daily company developments, track computing usage by product area and follow internal discussions, calling the setup a productivity gain that keeps humans in the loop on major decisions. He also said fully self-improving AI remains far off, a view echoed by Google DeepMind research showing AI writes over 75% of its production code but only 11.9% of code changes are fully delegated to machines, with no statistically significant acceleration in overall AI research pace.
Key Elements
Alphabet Inc. (GOOGL) Chief Executive Sundar Pichai says he now leans on a set of AI agents to keep tabs on the sprawling company he runs, a window into how far automation has crept into the daily work of even Google’s top executive.
Speaking on the Sources podcast in an episode released Wednesday, Pichai said simply staying current with developments across the company used to consume a substantial slice of his day. “It could easily take me an hour a day to kind of catch up, what’s happened in the last 24 hours,” he said. “But now I have agents running which summarize and bring that information to me. I can still go deeper, and I do so often.”
The agents serve several functions, according to the chief executive. They help him follow the arc of model development, track how computing resources are being consumed across product lines, and surface threads from internal chat groups without requiring him to read every conversation. He described agents with “telemetry into our compute usage by product area, efficiency, utilization, and giving me inputs and insights.”
“I think it’s made all of us, you know, be better at what we do,” he said.
Pichai was careful to note that people remain in the loop on decisions that matter, including how computing capacity is allocated. The agents, he said, are a tool for gathering and synthesizing information rather than a substitute for judgment.
He also pointed to a structural advantage of cloud-based agents: they keep working when he does not. “Because your agents, they’re running in the cloud, they’re running 24/7, they have your context, and so they can pick up things where you left off,” he said, describing a handoff in which the software continues on one task while he turns to the next.
The remarks land as Google leans heavily on AI to build its own systems. Pichai said in a separate interview published Wednesday that people remain deeply involved in that process. “There’s a lot of humans in the loop on these things,” he said, noting that models already propose training methods and evaluate results, with further acceleration expected over the next 12 months.
He has also sought to temper expectations about how quickly AI could improve itself without human direction. “We’re pretty far from … the end state of RSI,” Pichai said on the Big Technology Podcast, referring to recursive self-improvement, in which a system designs ever more capable successors on its own.
The company’s own research offers a mixed picture of where automation stands. A Google DeepMind safety report dated Oct. 13 found that AI writes more than 75% of the unit’s production code, yet only 11.9% of code-change submissions were fully delegated to AI in September, up from 7.5% in April. Researchers examined progress across 92 benchmarks and concluded that an apparent pickup in the pace of development was not statistically significant, finding “no significant end-to-end acceleration of AI R&D, AI-driven or otherwise.”
| Measure | Figure |
|---|---|
| DeepMind production code written by AI | More than 75% |
| Code-change submissions fully delegated to AI (September) | 11.9% |
| Code-change submissions fully delegated to AI (April) | 7.5% |
| Benchmarks tracked in DeepMind assessment | 92 |
| Gemini training-time reduction from AlphaEvolve optimization | About 1% |
Note: Figures come from Google DeepMind’s Oct. 13 safety report and Google’s disclosure of an AlphaEvolve optimization.
AI is nonetheless delivering concrete gains. Google’s AlphaEvolve system found an optimization that trimmed Gemini training time by roughly 1%, the company has said. The assessment found AI most useful for well-defined tasks, while researchers still lead on choosing which scientific ideas to pursue, designing experiments and resolving unexpected training problems.
Rivals are chasing similar ground. OpenAI said in September it had reached its automated research-intern milestone and aims for a more capable automated researcher by March 2028, still under human oversight. Anthropic says its Claude model writes more than 80% of the code merged into its codebase, while maintaining that people remain essential to picking worthwhile research problems.
Google co-founder Sergey Brin has taken a larger role in Gemini development as Anthropic and OpenAI have gained ground. He has said the company is “starting to get onto that RSI treadmill,” with researchers and models working together on new Gemini versions, and has pressed teams to move faster even without holding a formal executive title.
The push carries commercial weight. Pichai has acknowledged that Google sacrificed potential cloud revenue to prioritize AI research, even as Google Cloud revenue jumped 82% to $24.8 billion. He has called supplying computing power to Google DeepMind “non-negotiable” and warned that meeting customer demand for capacity in 2027 will be “incredibly difficult to do.”
For investors, the disclosures sketch a company betting that internal automation compounds into faster model development, while its own researchers caution that the productivity payoff has yet to show up clearly in aggregate. The gap between widespread AI assistance and true end-to-end acceleration is the number that matters as Alphabet tries to defend its position against OpenAI and Anthropic.
