Google New Gemini Skillz Turn Chrome Into an AI Beast

Key Concepts

• Skills in Chrome: A browser-level feature for saving and reusing AI prompt workflows.
• Agentic Workflows: Systems that move beyond simple chat to multi-step, goal-oriented execution.
• Gemini Robotics ER 1.6: An embodied reasoning model for robots, focusing on spatial awareness and instrument reading.
• Agentic Vision: A technique where AI zooms into images and runs code to interpret complex visual data (e.g., analog gauges).
• Vantage: A research framework using "Executive LLMs" to evaluate human soft skills like creativity and conflict resolution.
• Multimodal Reasoning: The ability of models to synthesize information across text, vision, and action.

1. Chrome "Skills" and Browser-Level Automation

Google has introduced "Skills" in Chrome, allowing users to save complex prompts as reusable workflows.

• Functionality: Users can trigger saved prompts with a single action (slash command or plus button) across multiple tabs simultaneously.
• Technical Shift: This moves prompt management from developer-side code (e.g., LangChain) to a user-facing UI.
• Safety: Includes "confirmation gates" for high-impact actions like sending emails or calendar invites to prevent unauthorized execution.
• Availability: Launched April 14, 2026, for Mac, Windows, and Chrome OS (English US).

2. Gemini Enterprise Agent Tab

Google is evolving Gemini from a chatbot into an execution workspace.

• Structure: The new "Agent Tab" features an "Inbox" and "New Task" interface.
• Workflow Management: Includes a side panel for defining goals, connecting apps, and managing files.
• Human-in-the-loop: A "require human review" toggle indicates a shift toward agents capable of desktop-level actions, potentially signaling a future standalone Gemini desktop application.

3. Notebook LM: Canvas and Connectors

Notebook LM is transitioning from a document summarizer to a research platform.

• Canvas: A visual layer that transforms text sources into timelines, interactive pages, or lightweight apps.
• Connectors: A new feature allowing the tool to pull data from external services, moving beyond manually uploaded files.
• Organization: Improved source management through Gemini-powered auto-labeling.

4. Gemini Robotics ER 1.6

DeepMind’s upgrade to its embodied reasoning model significantly improves robot autonomy.

• Architecture: Distinguishes between the VLA model (Vision-Language-Action, which controls movement) and the ER model (Embodied Reasoning, which acts as the strategist).
• Spatial Reasoning: Improved ability to point, count, and map object relationships, reducing hallucinations (e.g., grabbing non-existent objects).
• Instrument Reading: A breakthrough capability where robots (like Boston Dynamics' Spot) can read analog gauges and digital displays.
• Performance: Using "Agentic Vision," the model achieved a 93% success rate in instrument reading, up from 23% in version 1.5.

5. Vantage: Measuring Human Skills with LLMs

Google Research unveiled "Vantage," a system designed to quantify subjective human skills.

• Methodology: Uses an "Executive LLM" to control multiple AI personas in a conversation, steering the dialogue to test specific competencies like conflict resolution or project management.
• Evidence: In tests with 188 participants, the system achieved a 92.4% information rate for project management and an 85% rate for conflict resolution.
• Accuracy: AI scoring showed a Pearson correlation of 0.88 with human experts when evaluating creativity, demonstrating high reliability for subjective assessment.

Synthesis and Conclusion

Google’s recent updates represent a strategic shift toward agentic systems—moving from isolated AI interactions to persistent, multi-step workflows. Whether through browser-based automation (Chrome Skills), desktop-level task execution (Gemini Enterprise), or physical-world reasoning (Robotics ER 1.6), the common thread is the transition from "AI as a chatbot" to "AI as an operator." The introduction of Vantage further suggests that Google is applying these same agentic principles to measure and simulate human behavior, creating a closed-loop ecosystem where AI can plan, execute, and evaluate complex tasks across digital and physical environments.