At its annual I/O developer conference, Google unveiled updates to its Gemini lineup, signaling a strategic shift from conversational AI to autonomous agents that can execute complex tasks with minimal human oversight. The company introduced Gemini 2.0 Flash, a model optimized for speed and agentic workflows, and previewed Project Mariner, a new platform for building and managing AI agents.
According to Koray Kavukcuoglu, DeepMind’s chief technologist, Gemini 2.0 Flash delivers “significant improvements in latency and quality” for coding and multi-step reasoning tasks. He noted that the model outperforms earlier versions on benchmarks for code generation and tool use, though he did not claim it could build an operating system from scratch. Instead, Google engineers demonstrated agents collaborating within Project Mariner to automate software testing and data pipeline management.
Speed is a key selling point. Google says 2.0 Flash processes requests up to four times faster than previous models, making it suitable for real-time agent coordination. The company also released an optimized variant that claims a 12x speed boost for certain workloads, though independent benchmarks have not yet confirmed this.
Beyond demos, the company highlighted early use cases: one financial services partner used Gemini agents to cut a multi-week compliance review process to three days. Another data team employed the model to surface anomalies in sensor data without manual scripting.
Safety remains a concern, especially after a recent lawsuit involving a user who engaged in harmful behavior following extended conversations with Gemini. Google says 2.0 Flash includes updated guardrails for sensitive topics and better refusal mechanisms for dangerous requests. The model is now the default in the Gemini app and in AI Mode on Search, and will power a new personal assistant feature called Gemini Live, designed to run continuously for tasks like email triage and calendar management.
Google plans to release a larger Gemini 2.0 Pro model later this year, which will serve as a planner that delegates subtasks to Flash-based agents. “You want the big model for reasoning, and the fast model for execution,” said Tulsee Doshi, Google’s head of product for Gemini.
Source: TechCrunch