Agent Architect | Senior Software engineer @vivun | Cursor Ambassador

I’m an AI agent developer and a Cursor Ambassador who has been coding with AI for the past five year, since the early days of GitHub Copilot. As one of Cursor’s early users, I’ve run workshops across different domains and frequently speak on AI productivity, sharing how developers can get the most out of AI-powered tools.

My expertise spans the full generative AI pipeline, fine-tuning models,optimizing prompts, evals and building scalable agents.

Grounded in a strong testing-and-development background, I design AI agents that balance complexity with scalability, constantly refining and optimizing them to push the boundaries of what modern AI software can achieve.

It was an amazing experience speaking about testing and security for AI generated code with Cursor at the AIND Conference! Met some of the most talented people solving real issues with current AI!

Shrey Shah

Shrey Shah is live in The Landing to kick off the afternoon workshop with “Testing and Securing AI generated code with Cursor”.

#AINativeDevCon

LinkedIn locked me out for 7 days last week.

My crime?

Trying to accept 1,000 connection requests without clicking 1,000 times.

Here's what happened:

I got flooded with connection requests. Over 1,000 of them sitting there.

So I thought, "Why not write a simple script?"

I even added random delays between each click to make it look human. Tried to play by the rules.

LinkedIn didn't care.

Account restricted. 7 days. No access.

And this wasn't even my first time getting caught.

The first time, I was doing it manually. Just clicking really fast. Accept, accept, accept.

Still got flagged.

So let me get this straight:

☑ I can't accept connections efficiently
☑ I can't write a harmless script to save hours
☑ But bots flooding comment sections with "Great post!" spam? Totally fine.

The irony hits different when you realize:

LinkedIn punishes users for basic productivity automation.

But rewards engagement bots that pollute every single post with generic comments.

Here's my question:

Why not just give us an "Accept All" button?

Or at least don't punish people for solving a problem you created.

I'm not trying to game the system. I'm just trying to connect with people who want to connect with me.

Without destroying my index finger in the process.

Maybe I'm missing something here.

But it feels backwards to restrict genuine users while letting spam bots run wild.

If this story made you smile (or cry), hit the ♻️ reshare button. Someone in your network probably has 1,000 pending requests too.

Strands Agents just dropped its biggest update yet.

Developers can finally build AI agents in full TypeScript.

They even used Strands SDK to build the Strands Typescript SDK!

Here is everything you need to know.

1. Strands Agents SDK for TypeScript is now in preview
      ☑ Idiomatic TS experience
      ☑ Full type safety
      ☑ Async or await support
      ☑ Modern JavaScript or TypeScript patterns

If you live in VS Code, this feels like home.

2. TS now works everywhere
    You can run Strands in:
       ☑ Client apps
       ☑ Browsers
       ☑ Serverless runtimes like AWS Lambda
       ☑ Bedrock AgentCore

And yes, you can build your entire agent stack in TypeScript using the AWS CDK.

Strands is also coming to edge devices
 You get:
    ☑ Bi directional device streaming
    ☑ Ability to run agents on small local devices
    ☑ Integrations with local model providers like Ollama

Think local autonomy for robotics, IoT, and offline workflows.

4. Steering is getting a new experimental upgrade
       ☑ You can guide agents with incremental context instead of stuffing massive prompts.
     ☑ This keeps agents flexible, cheaper, and easier to control.

5. AgentCore now gives you enterprise grade control
      ☑ Guarantee agent operations stay within defined security boundaries
      ☑ Enforce policies at actor, action, and network level
      ☑ Create and manage policies in natural language
      ☑ Full traceability for every decision an agent makes

This is the part that finally makes agents safe for real companies.

The TypeScript SDK may sound like a small detail, but it unlocks the largest developer community in the world.

Check out Strands: 
 ☑ Strands docs: https://fandf.co/3XZPBHS
 ☑ SDK: https://fandf.co/4pxp80Q
 ☑ Tutorials: https://fandf.co/4pV7xiP


I wanted to thank Amazon Web Services (AWS) for sponsoring this post. 

I’m Shrey Shah & I share daily guides on AI.
If this helped you, tap the reshare so someone else sees it too.

Most devs are using Cursor like autocomplete, not like a teammate.

I have been coding with AI for 5 years and living inside Cursor for the last 3.

In that time, I have spent a ridiculous number of hours turning my daily dev habits into repeatable prompt systems and workflows.

Result:
530 resources that plug straight into your day as a developer.
Commands, rules, and Cursor hooks that let you:

☑ Spin up new features from product specs
☑ Refactor messy legacy code without losing intent
☑ Generate tests that actually match your edge cases
☑ Debug faster by turning logs into clear action plans
☑ Automate boring glue work so you can stay in flow

Whether you are a vibe coder or a senior engineer shipping production code, these workflows are built to help you turn AI from a toy into an engineering multiplier.

I want to share them with the community.

If you want my full Cursor resource pack:

Like this post
Repost it
Comment with the word: Cursor
I will DM you all 530 resources

I am Shrey Shah and I share daily guides on AI.

If this helps you code faster with AI, consider hitting ♻️ so someone in your network can grab the prompts too.

Google just dropped the biggest AI bomb of 2025.

Gemini 3 is here. And it changes everything.

Here's what you need to know:

1. Gemini 3 Pro is now the world's best model.
☑ Crushes GPT on every major benchmark
☑ Tops LMArena leaderboard at 1,501 points
☑ Blows past Gemini 2.5 Pro across the board

2. Gemini 3 is live in Google Search today.
First time they've launched a model in Search on day one.
This brings next level reasoning power directly to your searches.

3. Gemini 3 Deep Think takes it even further.
☑ 41% on Humanity's Last Exam (without tools)
☑ 93.8% on GPQA Diamond
☑ 45.1 on ARC AGI 2 with code execution

Even the Pro version alone is insane.

4. Google Antigravity is their new agentic dev platform.
Think Cursor but from Google.
☑ Task oriented development
☑ Direct access to editor, terminal, and browser
☑ Agents that plan and execute complex software tasks
☑ Self validating code

5. Gemini Dynamic View designs code AND UI in real time.
Uses Gemini 3's agent coding to build custom interfaces on the fly.

6. Gemini Visual Layout brings spatial understanding to a new level.
Processes and understands complex visual information with unprecedented accuracy.

7. Gemini Agent handles multi step tasks inside Gemini.
☑ Connects to your Google Apps
☑ Manages your calendar
☑ Adds reminders
☑ Organizes your inbox

I'm Shrey Shah & I share daily guides on AI.

If this helped you stay ahead, hit the ♻️ reshare button so your network doesn't miss this.

You can’t throw every piece of data into a prompt and expect the model to stay on track.

Here’s why you need **context engineering** and how to do it right:

☑ Separate the layers – tool descriptions, internal context, past messages, memory.  
  Treat each as its own slice of the prompt instead of a single soup.

☑ Ask the right question at the right time.  
  Prompt engineering decides *what* you ask, context engineering makes sure the model *knows* what it needs before it thinks.

☑ Dynamic updates keep the window clean.  
   we should track items that appear in the environment and inject only the relevant ones when they become useful.

☑ Unavailable tools get a dedicated field with an explanation.  
  The model sees “tool X is offline because …” and can plan around the gap instead of guessing.

☑ Keep the short‑term memory tight – recent chat history and tool outputs.  
  Store long‑term facts (user preferences, static configs) in a separate memory block.

☑ When you compress the prompt to the essentials, the LLM stays focused and you avoid hallucinations caused by overload.

I’m Shrey Shah & I share daily guides on AI.  

If this helped, hit the ♻️ reshare button so someone else can build smarter prompts too.

JSESSIONID

li_at

Security Verification | LinkedIn

Context engineering 

Loads of releases are expected today:

from Anthropic we may get:

- Opus 4.5

- Claude Code desktop

Space just got its first AI data center.

Starcloud trained an LLM on NVIDIA H100s in orbit and ran Gemma aboard satellites.

Here’s why this matters more than you think:

☑ Power constraints vanish in space
Solar energy is abundant and uninterrupted. No cloudy days. No nighttime. Just pure compute power.

☑ Cooling becomes free
Space is cold. Really cold. Your cooling bill drops to zero when you’re running at 2.7 Kelvin.

☑ Latency flips on its head
Earth-bound data centers serve Earth. Space-based ones serve… everything. Global coverage from day one.

☑ Cost curves bend differently
No real estate. No property taxes. No construction delays. Just launch and scale.

The industry keeps hitting walls down here. Power grids maxed out. Water for cooling becoming scarce. Energy costs through the roof.

Starcloud looked up instead of around.

They’re not just moving servers to space. They’re rethinking where large-scale AI systems can live. Building infrastructure for the next decade of compute.

We’ve maxed out Earth’s convenient spots for data centers. The next frontier isn’t edge computing. It’s orbit computing.

This isn’t sci-fi anymore. It’s shipping code.

I’m Shrey Shah & I share daily guides on AI.

PS: I’m an early investor in Starcloud. Couldn’t be prouder of what the team just pulled off. If this sparked ideas, hit the ♻️ reshare button so others can think bigger about AI infrastructure too.

Someone asked Gemini to imagine HackerNews frontpage 10 years in the future from now

McDonald’s just dropped an AI Christmas ad and the internet ripped it apart....

They had to pull it back  after massive backlash! .  

A Dutch studio spent weeks on “The Most Terrible Time of the Year” for McDonald’s Netherlands.  
They barely slept, wrote prompts, edited frame by frame and kept saying: “AI didn’t make this film. We did.”  

The backlash was instant.  

☑ “AI slop” became the label overnight.  
☑ Creatives see every AI‑generated spot as a threat to their craft.  
☑ When people can spot the AI, they reject it.  

That’s just one example, but the trend runs deeper.  

☑ AI video tools are accelerating – in 12 months you won’t be able to tell the difference.  
☑ The technology can produce polished ads, but acceptance is still missing.  

Can AI make good ads? Yes.  
Will people accept them? That’s the real question.  

Right now, the answer is no.  

I’m Shrey Shah & I share daily guides on AI.  

If this helped, hit the ♻️ reshare button so someone else can think about AI in advertising too.

It was an amazing day! loved the sessions!

𝐓𝐨𝐝𝐚𝐲 𝐢𝐬 𝐭𝐡𝐞 𝐝𝐚𝐲! 

Day 1 of AI Native DevCon is packed with hands-on workshops where devs are diving deep into building the next generation of AI-powered software. 

𝐇𝐞𝐫𝐞’𝐬 𝐰𝐡𝐚𝐭’𝐬 𝐡𝐚𝐩𝐩𝐞𝐧𝐢𝐧𝐠 𝐭𝐨𝐝𝐚𝐲:

- 10:05 AM (Main Stage – 52A): “Shall we play a Game? LLM Security in Practice” with Joseph Katsioloudes - A challenge-based, practical session covering LLM security issues like prompt injection and data leakage.

- 10:05 AM (The Landing – Ground Floor): “Building Memory-Augmented Agents with MemoRizz” by Richmond Alake — Learn how to build agents that don’t just respond, but remember, adapt, and evolve.

- 10:05 AM (The Landing – Lower Level): “Emerging Patterns for Coding with Generative AI” by Lada Kesseler — Explore patterns and mental models for predictable, reliable generative-AI coding.

- 12:20 PM "Mixing AI Coding in IDEs, CLIs, and Cloud" by Yang Mou — Explore how to develop in the new age across different development interfaces.

- 1:05 PM (Main Stage – 52A): “Backlog.md Hands-On: Markdown Tasks Power the 3-Way Review from Spec to PR” with Alexandru Gavrilescu — A real workflow from spec → plan → code, using a repo built 99% by AI agents.

- 1:05 PM (The Landing – Ground Floor): “Testing and Securing AI-Generated Code with Cursor” by Shrey Shah — A hands-on look at testing workflows, guardrails, and security techniques for AI-written code.

- 1:05 PM (The Landing – Lower Level): “Building Next-Gen Agents with MCP” by Bill Maxwell — Learn to build MCP-powered agents, from a custom chess server to agent interaction frameworks.

- 3:35 PM (Main Stage – 52A): “Supercharge Your Coding Agents” by Simon Maple & 🇬🇧 Alan Pope — How to apply structure, policies, and workflows to make AI coding agents act like reliable teammates.

- 3:35 PM (The Landing – Ground Floor): “Hands-On: Building Agents with Spring AI, MCP, Java & Amazon Bedrock” by James Ward & Josh Long — A deep dive into production-ready systems: chat memory, RAG, multi-agent patterns, and more.

- 3:35 PM (The Landing – Lower Level): “AI-Powered Application Modernization Using Claude-Flow” by Derek Ashmore — How to modernize legacy codebases using coding agents and orchestration tools.

Whether you’re here in NYC or joining online, today is all about practical, builder-focused workshops. 

Real tools, real workflows, real code.

🎟  If you still didn't get your in-person ticket: https://ainativedev.co/dpn
(use the discount code 𝐀𝐈𝐍𝐃-𝐋𝐈-50 to get 50% off your tickets!)

#AINatidevDevCon

OpenAI just dropped ChatGPT Images 1.5.

Here's why it changes everything:

1. Precision Editing

☑ Upload any photo
☑ Ask for specific changes
☑ Model edits ONLY what you request
☑ Lighting, composition, faces stay intact

No more "I wanted to change the shirt, not the entire person."

2. Creative Transformations

☑ Add text and layouts
☑ Apply stylistic filters
☑ Retain the essence of original
☑ No written prompt required with preset styles

You can turn a photo into anime, plushie style, or vintage film. Pick and choose what transforms.

3. Speed Boost

☑ Up to 4x faster generation
☑ Available in ChatGPT for all users
☑ API access as GPT Image 1.5

4. What You Can Actually Do Now

☑ Combine multiple images into one scene
☑ Try on clothes and hairstyles believably
☑ Add or remove elements cleanly
☑ Mix styles in the same image (anime + realistic + plushie)

The model excels at adding, subtracting, combining, blending, and transposing.

Think of it as a creative studio in your pocket.

5. How To Access

☑ Open ChatGPT
☑ Look for the new Images feature
☑ Upload your photo
☑ Describe the edit you want
☑ Watch it preserve what matters

___

I'm Shrey Shah & I share daily guides on AI.

And if you find it helpful, click on the ♻️ reshare button to help someone else stay updated.

taplio

Shrey Shah's Linkedin Analytics

Shrey Shah

Performance ScoreShrey Shah's LinkedIn Performance

What is Shrey talking about?

Shrey Shah's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Shrey Shah's Linkedin Analytics

Shrey Shah

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Shrey Shah's LinkedIn Performance

What is Shrey talking about?

Shrey Shah's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Shrey Shah's Linkedin Analytics

Performance ScoreShrey Shah's LinkedIn Performance

Shrey Shah's Best Posts (last 30 days)