- Building AI Agents
- Posts
- Gemini 3 leaves the competition in the dust
Gemini 3 leaves the competition in the dust
Plus: Google’s online agent building course, Microsoft shifts from “per user” to “per agent” pricing, and more
Michael Cunningham
November 20, 2025
Edition 138 | November 20, 2025
To our Albanian readers: is…everything okay? Blink twice if you need us to call the UN.
Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!
In today’s issue…
Big model releases by Google and xAI
Google’s online agent building course
OpenAI’s cookbook for context engineering
Microsoft shifts from “per user” to “per agent” pricing
The 5 core types of agent failures
…and more
🔥 IN CASE YOU MISSED IT
Readers’ favorite items from the past week
‼️ ESSENTIAL NEWS
Source: Google
The company’s new flagship model was launched on Tuesday, achieving very impressive leaps over the previous state-of-the-art models on nearly every benchmark, including extremely difficult ones like Humanity’s Last Exam and ARC-AGI-2. Its Deep Think mode—not available yet while it undergoes safety evaluations—scored even higher. The non-Deep Think version of Gemini 3 Pro is now available to agent builders in the Gemini API.
Along with the model itself, Google also released Antigravity, an agentic software development tool powered by Gemini 3 Pro that allows users to run several coding agents in parallel, a style that many power users have been trending towards but which is not easily supported by other agentic coding platforms.
Elon Musk’s xAI unveiled the latest of its Grok models, which claimed the #1 spot on the LLM Arena leaderboard until it was de-throned a day later by Gemini 3 Pro. Along with the new model, xAI released the Agent Tools API providing Grok with a set of capabilities that other major LLM providers have already incorporated, such as web and file search, code execution and MCP integration.
🤖 AGENT OF THE WEEK
👋 Welcome back to Agent of the Week!
The meeting is never the time consuming part. The follow-up is.
I used to finish a great call, promise the world "I'll send that contract," "Let's meet next Tuesday" and then immediately get distracted.
The notes stayed buried. The momentum died. The email got sent three days late.
So I built an agent to fix it.
Meet Exec, my meeting follow-up agent.
Exec connects to my Fathom transcripts, but it doesn't just summarize the meeting. It actions on it.
It automatically distinguishes between a task ("I'll fix the bug"), a meeting ("Let's sync Tuesday"), and a promise ("I'll email you").
Now, when I open my laptop:
The calendar invite is already created.
The tasks are logged in Sheets.
The follow-up email is written and waiting in my Drafts folder.
And the best part? Since the agent is built modularly, you can always let it do more than just these 3 things.
Let me show you how I built this one.
🧠 How It Works
1. Webhook Trigger → Meeting Completed: As soon as the Zoom call ends, Fathom AI sends the full transcript to the agent — this works with any meeting
2. AI "The Brain" → Structured Extraction: The agent reads the transcript and separates conversation into three distinct buckets: Tasks, Meetings, and Emails. It even calculates the correct dates based on when the meeting happened
3. The Router → Traffic Control: The agent splits the data and routes it to the correct tool in my Google Workspace
4. Google Sheets → Task Logging: If I said "I will fix the slide deck," Exec logs it as a row in my Master Task Tracker
5. Google Calendar → Auto-Scheduling: If I said "Let's chat next Tuesday at 2pm," Exec handles the timezone math and creates a calendar event automatically
6. Gmail → The "Draft" Magic: If I said "I'll send you the pricing PDF," Exec writes the email for me, addresses it to the client, and saves it as a Draft. I just hit send
The Big Shift: Most people use AI just to chat. That gives you words, but you still have to do the clicking.
The real value happens when you give the AI tools. When you connect the Brain (LLM) to the Tools (Gmail, Calendar), you stop playing with a chatbot and start building a digital assistant.
Want to create your own? I dropped the full copy/paste template & prompt inside the Building AI Agents Community. Come create your own agent!
Till next week,
✌️ AP