Google Chief Scientist Jeff Dean said in an interview with Y Combinator that AI agents are evolving from tools into systems capable of running continuously for weeks, with the core bottleneck shifting from model capability to inference hardware efficiency. Dean stated his May 2025 prediction that "AI has reached the level of a junior engineer" has largely been validated, and that progress on complex tasks has been faster than he expected. He emphasized that agent-based systems are showing capabilities beyond programming, and can run continuously for days or even weeks given certain problem domains and sufficiently powerful models. Dean proposed that reducing inference latency by 50 times would fundamentally alter the product boundaries of AI systems and give rise to entirely new application paradigms. He signaled to entrepreneurs that context engineering, domain-specific models, and multi-agent systems capable of integrating inference-time compute will be primary windows for small teams to achieve differentiation. Dean cited an example of teaching a model to complete a workflow of "measure baseline—improve code—measure performance gain—iterate" while optimizing low-level library performance with his colleague Sanjay, and noted models are already efficient at translating software between programming languages.