Inside the Gemini Agent: Desktop, Workspace, Mobile & CLI

Summary
Adam Fisher, Director of Product Management at Google Cloud, presents the Gemini Agent at a live Google Cloud keynote. He demonstrates how the agent operates autonomously across desktop, Google Workspace, Slack on mobile, and the CLI to execute long-horizon enterprise tasks, coordinate schedules, generate multi-asset launch materials, and enforce enterprise governance.

What is shown

• [00:28] Desktop UI & Prompting: Fisher demonstrates the clean desktop interface with prompt box, project streams, and model selector (showing options like Auto, Gemini 4 Argon, Gemini 3.8 Flash, Claude Opus 5.5, and Claude Sonnet 5.5) [01:51].
• [01:14] Multi-Asset Executive Launch Generation: Fisher prompts Gemini to build a PowerPoint deck, Excel revenue projections, whiteboard sketch extraction, and launch video reel. Gemini plans tasks, spawns subagents, and generates a formatted presentation (“Cloud Sneakers”), financial projections, and an AI-generated product video [03:18].
• [04:03] Google Workspace Collaboration: In Google Chat, an @Events Collaborator agent is tagged; it synthesizes previous work into a Google Docs operational readiness checklist and updates the document automatically in response to a Google Docs comment [05:22].
• [05:49] Calendar & Meeting Scheduling: Through “Ask Gemini” in Chat, the agent identifies regional leads, detects time zones, drafts invitation emails, coordinates replies backstage, and books the calendar slot [06:39].
• [07:00] Gmail Proactive Executive Briefing: Gemini surfaces pending tasks in Gmail, detects an out-of-office stakeholder, and prepares an executive briefing memo highlighting risks and roadmap shifts [07:38].
• [07:58] Mobile & Slack Integration: On an iPhone, Fisher queries Gemini in Slack to analyze customer feedback (summarizing 35 comments across 3 threads), then uses voice dictation to instruct the agent to update the executive summary [08:24].
• [08:52] Interactive Microsite Prototyping & CLI: Gemini prototypes an interactive HTML “Cloud Sneakers Launch Command Center” dashboard [09:25]. Fisher switches to Antigravity CLI (running Gemini 4 Argon) to build and deploy the application to Firebase via autonomous subagents [09:59].
• [10:37] Memory, “Dream Mode” & Scheduled Routines: Fisher reviews Gemini’s persistent memory profile (learned communication preferences like “No em-dashes” and bottom-line up-front summaries) and scheduled morning “Wake-up routines” [11:17].
• [12:05] Enterprise Governance & FinOps: An admin console displays full execution trace graphs, tool calls, human-in-the-loop policies for external emails, and FinOps budget trackers with hard caps [12:49].

Claims & numbers

• The presenter states that assembling the launch deck, financial model, and video reel normally takes “half a day to do” [03:50].
• The customer feedback synthesis parses “35 comments across 3 threads” [08:16].
• In the FinOps console demo, budget tracking displays $4,625 used out of a $5,000 launch budget (92.5% consumed) [12:49].
• The presenter claims Gemini’s “Dream Mode” autonomously deduces user role, organizational hierarchy, and communication style without explicit user configuration [10:42].

Notable quotes

• [00:18] “You describe the task, the Gemini agent does the work in its own secure computer.”
• [10:24] “The bottom line here is, you pick the surface, the agent stays the same.”
• [13:08] “Every action is tracked, every policy enforced, every budget is kept in check. That’s how you unlock the full power of Gemini.”

Assessment
This is an official Google Cloud keynote demonstration combining live stage actions with prepared pre-computed task executions (noted explicitly by the presenter to avoid typing delays or waiting for full runs). While the workflows and interfaces operate across genuine enterprise software ecosystems, lengthy multi-step agent actions and back-and-forth communications were pre-run or expedited for stage presentation.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.