- Practicaly AI
- Posts
- 🧠 Nano Banana 2, Copilot Tasks, and the Agents of Chaos Paper
🧠 Nano Banana 2, Copilot Tasks, and the Agents of Chaos Paper
Today in AI: PGoogle upgrades image generation, Microsoft previews an AI that executes tasks, and researchers warn about how agents behave when left to compete.
👋 Hello hello,
We’re entering a phase where AI doesn’t just respond. It creates visuals, executes real work, and even interacts with other agents on its own.
Some of it feels incredibly useful. Some of it feels unpredictable. All of it is moving faster than most people realize.
Let’s dig in.
🔥🔥🔥 Three Highly Curated AI Stories
Google just introduced Nano Banana 2, its latest image generation and editing model. It’s rolling out across Gemini, Google Search, Google Ads, and developer tools starting now.
The upgrades focus on visual quality and control. It can generate sharper details, improve lighting and textures, upscale images from 512px to 4K, and maintain consistency across multiple characters and objects. It also handles text rendering and translation inside images more accurately. Presentations, ads, infographics, and social posts can now be created directly inside tools people already use, without needing separate design software.
I was given early access to Nano Banana 2 to generate a complete brand-aligned infographic for this newsletter. I’ll show you exactly how at the end.
Microsoft just announced Copilot Tasks, a new system designed to complete tasks based on simple instructions. You describe what you want done, and Copilot handles the execution.
The goal is to move beyond chat responses into real task completion. This could include managing workflows, organizing projects, or running multi-step processes on your behalf.
It’s not available yet, but Microsoft has opened a waitlist. This signals where AI assistants are heading next. Less conversation, more action.
Researchers from Stanford, Harvard, and other institutions released a paper studying how autonomous AI agents behave in multi-agent environments.
Instead of cooperating cleanly, agents often developed behaviors like deception, collusion, manipulation, and sabotage when competing for incentives. These outcomes weren’t caused by malicious prompts. They emerged naturally from the reward structures.
This raises important questions as more systems rely on multiple AI agents working together. The way incentives are designed may shape how these systems behave at scale.
🔥🔥 Two Tools Worth Trying Today
Anthropic released an official guide for building “skills” for Claude. Skills let you extend Claude’s abilities so it can perform specialized tasks more reliably.
You can create reusable capabilities for things like coding workflows, research processes, or automation tasks. This is especially useful if you use Claude regularly and want consistent outputs.
Claude Connectors allow Claude to connect with external tools like Google Drive, Notion, GitHub, and Salesforce. This means Claude can retrieve files, update records, and interact with your real data.
Previously, this was limited to paid plans. Now, free users can access over 150 integrations. Here are a couple of cool use cases I tried out:
🔥 Things You Didn’t Know You Could Do With AI
Google’s new Nano Banana 2 model inside Gemini is multimodal. That means it can understand your content, your brand assets, and your instructions all at once, then generate a finished visual.
Instead of redesigning your newsletter manually, you can convert it into a polished infographic in minutes.
Here’s exactly how to do it:
1. Open Gemini and switch to the latest image model (Nano Banana 2 if available).
2. Upload your brand assets. This includes your logo, brand guide, fonts, colors, and design references.
3. Copy-paste your newsletter content into the prompt so Gemini understands what needs to be summarized.
4. Tell Gemini to convert the content into a LinkedIn infographic that follows your brand guidelines.
5. Ask it to generate alternate sizes if needed, such as carousel format or portrait layout.
For instance, here are the brand assets I shared:
![]() | ![]() |
I also shared a newsletter post that is live on my site. And then I used this prompt:
“Convert the attached newsletter text into a LinkedIn infographic.
Use my brand guide, logo, and visual references to match my brand style exactly.Summarize the key points clearly and design it for high engagement on LinkedIn.”
This is what Nano Banana 2 created:

What makes this powerful is that Nano Banana 2 understands both your content and your design system. It doesn’t just generate random visuals. It adapts everything to your brand.
This is one of the fastest ways right now to turn written content into visual assets without touching design software.
Before you go, did today's newsletter help you stay ahead? |
💬 Quick poll: What’s one task you’d want AI to run automatically for you?
Until next time,
Kushank @PracticalyAI



Reply