MOUNTAIN VIEW, CA — Google on Tuesday rolled out Gemini 2.5, a flagship update introducing what the company is calling “specialized agent architecture,” a suite of purpose-built AI models that includes one agent dedicated to scheduling, one to research, one to coding, and at least one whose sole responsibility appears to be coordinating the other agents because they have apparently stopped speaking to each other.
In a launch keynote streamed from a glass-walled room with the acoustics of a high-end yogurt commercial, Google DeepMind executives demonstrated how Gemini 2.5’s agents can independently book a flight, draft a follow-up email, and rewrite the follow-up email after a second agent flagged it as “too curt for Q2.”
“What we’re really excited about is the orchestration layer,” said Anjali Pradhan, a Google product lead whose title fit on the badge only in eight-point type. “Previous models tried to do everything in one context window, which created friction. Now each agent has a defined role, and a separate agent ensures they collaborate, and a third agent monitors that collaboration for compliance, and a fourth summarizes the whole exchange into a paragraph the user can ignore.”
Early benchmarks confirm Gemini 2.5 is the most capable model Google has ever shipped, scoring above competitors on reasoning, coding, and the increasingly important benchmark of “convincingly pretending to have read the attachment.” The company said the planning agent alone can break a complex task into as many as 240 subtasks, 239 of which are scheduling a meeting to discuss the first one.
Asked about energy usage, Google noted that Gemini 2.5 is “meaningfully more efficient per token” than its predecessor, a metric that becomes less reassuring once one learns the model is now generating roughly seventy times more tokens per query, most of them produced by agents asking each other to clarify earlier agents. A company sustainability report described the net environmental impact as “directionally encouraging,” a phrase the report’s authors used four times.
The launch arrives during a stretch in which every major tech company has pivoted to “agents” as the next inevitable computing paradigm, on the theory that consumers, having mastered typing things into a box, are now ready for a box that types things into other boxes on their behalf. OpenAI, Anthropic, and Meta are all expected to ship comparable agent suites within weeks, after which the agents will presumably begin negotiating with each other across platforms, generating a global volume of synthetic email that will require its own agent to triage.
Reaction among developers was cautiously enthusiastic. “It’s genuinely impressive,” said Marcus Thiele, a senior engineer at a mid-sized fintech who spent Tuesday afternoon watching a Gemini coding agent debug a problem caused by a Gemini deployment agent. “Although at some point I did notice I was just sitting there, and the agents were having a whole work life without me, and I started to wonder what the company was paying me for. Then an agent emailed to schedule a one-on-one about it.”
Google declined to specify the launch’s water and electricity footprint, citing competitive sensitivity, but a spokesperson confirmed the company remains “on track” to meet its 2030 net-zero targets, which were set in 2021, before anyone at Google had heard the word “agentic.” The same spokesperson clarified that the targets had not been revised, only “recontextualized.”
At press time, the orchestration agent had successfully delegated the writing of this article’s final sentence to four other agents, none of whom were responding.