Build ANYTHING with Gemini 3 | The Agent Factory Podcast

Key Concepts

• Gemini 3 Pro: Google's new flagship AI model, excelling in advanced reasoning, complex task completion, and agentic operations.
• Anti-gravity: Google's agentic coding environment, featuring numerous exploration-ready features.
• AI Studio: A platform for building AI applications, demonstrated for creating a personal portfolio website.
• Nano Banana: An AI image generator used for creating inspiration images and character visuals.
• Gemini CLI: A command-line interface tool that integrates with Gemini models, particularly useful for scaling tasks and automating workflows.
• AI Employees: AI agents designed to perform specific tasks within a business or personal workflow, aiming to increase efficiency.
• Standard Operating Procedures (SOPs): Markdown files used to define instructions, inputs, tasks, and desired outputs for AI agents.
• Agent Development Kit (ADK): Google's framework for building AI agents, enabling orchestration of multiple sub-agents.
• Script Sequencer: A sub-agent within ADK responsible for adapting scripts for natural delivery and chunking them into short, engaging segments.
• Video Agent: A sub-agent within ADK that generates video content based on scripts and character visuals.
• Human in the Loop: The concept of incorporating human oversight and intervention in AI-driven processes, especially for quality control and refinement.

Gemini 3 Pro and AI Studio Demo: Personal Website Creation

This section details the use of Gemini 3 Pro within Google AI Studio to build a personal portfolio website.

• Input Data:
  • LinkedIn profile exported as a PDF.
  • An inspiration image (inspo.png) generated by Nano Banana.
  • A profile image of the user.
• Process:
  1. The user uploads the LinkedIn PDF, profile image, and inspiration image into AI Studio.
  2. A prompt is given to Gemini 3 Pro: "Help me create a personal portfolio website based on the profile. PDF file which contains my LinkedIn profile and then use the inspo.png as inspiration for how the website should look like. Then make sure that you include that profile picture of me as well."
• Key Observations and Features:
  • Speed: Gemini 3 Pro demonstrated remarkable speed in generating the website, even under heavy load.
  • Advanced Reasoning: The model successfully extracted relevant information from the LinkedIn profile, including work experience, skills (e.g., "great public speaker, de machine learning"), and contact details.
  • Inspiration Integration: The inspo.png was used effectively to guide the website's visual style.
  • Annotation Feature: AI Studio's annotation feature was used to identify and correct specific issues, such as a missing or incorrect profile image. The user could select a section and instruct the AI to "add that section to chat and fix my image with the actual profile picture."
  • Single-Shot Capability: Gemini 3 Pro showed impressive performance on complex tasks with a single prompt, tested across various domains including web applications and advanced academic subjects like Calculus 3.
  • Consistency: The model maintained the initial black-and-white styling from the inspiration image and applied changes only to the specified section, avoiding unintended alterations to the entire website.
  • Deployment: The generated website can be easily deployed to Cloud Run with a single click, requiring no prior knowledge of Google Cloud or Cloud Run. This significantly lowers the barrier to entry for developers and individuals with ideas but no coding experience.

Gemini CLI and AI Employees: Market Research Automation

This section focuses on Brandon Hancock's use of Gemini CLI to build "AI employees" for market research.

• Goal: To automate tasks and free up time for higher-level strategic work by creating AI agents for specific business functions.
• Use Case: Building a market research AI employee for a startup that helps emergency medical services (EMS) with SOAP reports. The goal is to identify potential customers across America.
• Gemini CLI Features Highlighted:
  • Google Search Integration: The CLI can leverage Google Search to gather information at scale.
  • Model Variety: Supports multiple Gemini models, including Gemini 3 Pro for complex reasoning and Gemini 2.5 Flash for efficient, large-scale tasks.
  • Generous Free Tier: Offers 1,000 requests per day for free.
  • Markdown-Based SOPs: AI employees are defined by markdown files outlining Standard Operating Procedures.
• Building an AI Employee (Methodology):
  1. Define SOPs: Create markdown files that clearly specify:
     • Inputs: What data the agent will receive.
     • Tasks: The actions the agent needs to perform.
     • Outputs: The desired outcome of the task.
  2. Leverage Gemini Models:
     • Gemini 3 Pro: Used for writing well-documented, specific instructions for tasks like internet searches, defining the type of questions to ask, and the ultimate goal.
     • Gemini 2.5 Flash: Used for executing these instructions efficiently and affordably, especially for large-scale operations.
  3. Parallel Processing: Gemini CLI allows spinning up hundreds of agents in parallel to perform tasks concurrently.
  4. AI-Driven Script Generation: Brandon demonstrated that he doesn't manually code scripts. Instead, he prompts Gemini CLI with his requirements (e.g., "I want the ability to give you three or four cities to go off and research and what I want you to do is run you know Gemini-P as a prompt and basically just paste in this type of uh city right here or these type of instructions which is hey go research a city do it for this city uh and basically just kick this off 10 times or 10 20 times, however many cities I give you"). Gemini CLI then generates the necessary Python scripts to execute these tasks in parallel.
  5. Workflow Example (Market Research):
     • Research SOP: Instructions to research specific cities for potential EMS customers.
     • Python Script Execution: Gemini CLI, guided by the SOP, kicks off multiple parallel agents to research each city.
     • Data Aggregation: Information from different cities is aggregated into a database.
     • Outreach SOP: A subsequent SOP defines how to write personalized outreach messages to potential customers based on the aggregated research.
• Other AI Employee Applications:
  • Ghostwriter: Responding to emails by hooking up an IMAP server to Gmail. The AI knows the user's writing style and can draft responses. Human review is still required.
  • Report Generation: Automating the writing of monthly status reports.
• Core Principle: "If I'm ever doing something more than once, I can guarantee you AI is doing it for me."

ADK Agent: Educational Video Creation with Capybara Narrator

This section showcases an agent built using Google's Agent Development Kit (ADK) for creating educational videos.

• Agent Purpose: To create engaging educational videos from developer documentation.
• Key Technologies:
  • ADK: For orchestrating multiple sub-agents.
  • Gemini 3.1: The underlying AI model.
  • Nano Banana: For generating character visuals.
• Agent Architecture (Orchestrator and Sub-Agents):
  • Boss Orchestrator: Manages the workflow and interacts with sub-agents.
  • Script Sequencer:
    • Adaptation: Makes the script sound natural and engaging, adapting the text to the chosen character's persona (e.g., a Capybara). It avoids reading bullet points verbatim.
    • Chunking: Splits sentences into approximately 8-second segments, suitable for video generation models and ensuring natural transitions.
  • Video Agent: Generates video clips based on the sequenced script chunks.
• Video Generation Process:
  1. Input: A documentation page and four generated views of the Capybara character from Nano Banana.
  2. Script Adaptation and Chunking: The Script Sequencer processes the documentation, making it conversational and splitting it into short segments.
  3. Random View Selection: The Orchestrator randomly selects one of the four Capybara views for each video chunk to maintain visual engagement and avoid a monotonic talking head.
  4. Prompt Engineering for Consistency: Detailed character and voice descriptions are provided in prompts to ensure consistency in appearance (clothing, body parts) and environment across different video segments, even though video models often lack inherent context of previous generations.
  5. Video Generation: The Video Agent creates individual video clips for each script chunk.
  6. Output: A list of URLs for 92 generated video segments.
• Human in the Loop for Refinement:
  • The generated videos are not always perfect. The user manually reviews each 8-second segment.
  • If a video has glitches or sounds awkward, it can be regenerated. The agent can even be prompted to use a different view or adjust parameters.
• Video Assembly:
  • Gemini CLI: Used to join the individual video segments into a complete video. The user provides a list of video files and a command (e.g., using ffmpeg via Gemini CLI) to splice them together.
• Anti-gravity Integration for Agent Improvement:
  • Brandon demonstrated using Anti-gravity to modify the ADK agent. He opened a previous version and instructed Anti-gravity on desired changes (e.g., adding a "human in the loop" moment for script edits, adjusting prompts). Anti-gravity automatically updated the prompts and agent code.
• Resulting Video Demo: A short clip of the Capybara narrator discussing AI agent safety and security was played, highlighting its engaging and informative nature. The full 11+ minute video is available on GitHub.

Synthesis and Conclusion

The episode highlights the rapid advancements in AI, particularly with the release of Gemini 3 Pro and Anti-gravity. The discussion emphasizes the practical application of these tools for developers and businesses.

• Gemini 3 Pro is presented as a powerful model for building sophisticated AI agents, capable of advanced reasoning and complex task execution. Its integration into AI Studio simplifies the creation of applications like personal websites, lowering the barrier to entry for users without extensive coding knowledge.
• Gemini CLI is showcased as a crucial tool for scaling AI operations and automating workflows. The concept of "AI Employees" built using Standard Operating Procedures (SOPs) in markdown files, combined with the ability to generate and execute scripts in parallel, offers a powerful paradigm for business efficiency. The ability of Gemini CLI to generate its own scripts based on natural language instructions is a significant time-saver for developers.
• The ADK agent example demonstrates how to orchestrate multiple AI agents to tackle complex creative tasks, such as generating educational videos. The detailed breakdown of the Script Sequencer and Video Agent, along with the emphasis on character consistency through prompt engineering and the integration of Nano Banana for visuals, illustrates a sophisticated approach to AI-powered content creation. The necessity of a human in the loop for quality assurance and refinement is also acknowledged.
• The overarching theme is the increasing accessibility and power of AI tools, enabling individuals and businesses to automate tasks, accelerate development, and innovate at an unprecedented pace. The future vision includes agents interacting with each other, further streamlining processes.