Cloudflare Releases AI Agents SDK

Today’s top AI Highlights:

1. OpenAI GPT-4.5 API costs 25x more than Claude 3.7 Sonnet

2. Build stateful AI agents with Cloudflare’s new agents-sdk

3. TypeScript framework for agents and AI workflows

4. The first diffusion-based LLM for coding which is 5-10x faster

5. Opensource IDE extension to create, share, and use custom AI code assistants

& so much more!

Read time: 3 mins

AI Tutorials

Finding the perfect property involves sifting through countless listings across multiple websites, analyzing location trends, and making informed investment decisions. For developers and real estate professionals, automating this process can save hours of manual work while providing deeper market insights.

In this tutorial, we'll build an AI Real Estate Agent that automates property search and market analysis. It helps users find properties matching their criteria while providing detailed location trends and investment recommendations. This agent streamlines the property search process by combining data from multiple real estate websites and offering intelligent analysis.

Tech Stack:

• Firecrawl's Extract Endpoint to collect structured data from websites

• Agno (formerly Phidata) for building the AI agent

• OpenAI GPT-4o as the LLM

• Streamlit for a clean, interactive web interface.

We share hands-on tutorials like this 2-3 times a week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.

Don’t forget to share this newsletter on your social channels and tag Unwind AI (X, LinkedIn, Threads, Facebook) to support us!

Latest Developments

Build Agents, RAG, and AI Apps with React-like Components ⧉🔗

GenSX is a simple Typescript framework for building agents and workflows with reusable React-like components. It’s a workflow engine designed for building agents, chatbot APIs, and more using common patterns like RAG and reflection.

Unlike most frameworks that rely on cumbersome graph-based approaches with nodes, edges, and global state, GenSX uses JSX to define the flow of data. This provides a visually clear, top-to-bottom structure, eliminating the mental overhead of graph APIs and making it much easier to build, refactor, and evolve your AI applications.

Key Highlights:

1. Component-Based, Not Graph-Based - GenSX moves away from the typical node-and-edge graph approach. It uses pure TypeScript functions as components, making them inherently testable, reusable, and easy to share.

2. JSX for Data Orchestration - It uses JSX to define how data moves between your components. This provides a visually clear, top-to-bottom readable structure, unlike complex graph definitions, making dependencies easy to understand, and is something React devs are already familiar with.

3. Automatic Parallel Execution - Components execute in parallel when possible while maintaining dependencies, optimizing workflow performance without explicit configuration. The framework handles execution order and data flow automatically.

4. Built-in Streaming and Reflection - Toggle streaming with a single prop change without refactoring your code. The framework includes patterns for self-reflection, allowing LLMs to evaluate and improve their own outputs through recursive evaluation loops.

AI Search Just Had a DeepSeek Moment: Top AI Search Scores at 3% of Perplexity's Funding 🤯

The AI search space has its own DeepSeek moment. AI company Liner now tops OpenAI's SimpleQA benchmark, beating Perplexity Pro by 3.1 points while spending just a fraction of the capital.

Linear is an AI search engine that delivers factually accurate answers, helping you discover reliable sources, generate citations, and accelerate your learning.

How did Liner do it?

📌 Advanced Ranker Model – Trained on proprietary highlight data since 2015 to prioritize only the most reliable sources

📌 Intent-Driven Search LLM – Custom-built on Llama 3.3 to break complex queries into precise subquestions for superior accuracy

📌 Reference Utilization – Overcomes common LLM limitations with superior data grounding for accurate source citations

📌 Agentic Architecture – Orchestrates multiple specialized LLMs for different tasks from reasoning to fact verification

Want proof? Check out the live comparisons – it's honestly shocking how Perplexity Pro got basic facts wrong about Kara Walker's first exhibition and even when the Hellschreiber was invented. Meanwhile, Liner nailed every single one.

Try Liner Pro free for 2 weeks and experience what 11 million users already know – AI search can be both accurate and affordable.

Build and Deploy AI Agents on Cloudflare ☁️🤖

Cloudflare has released agents-sdk, a new JavaScript framework for building and deploying AI agents directly on its Workers platform. This SDK lets you create agents—autonomous programs that can perform tasks—and run them with the scalability and cost-effectiveness of Cloudflare's serverless environment.

The framework provides built-in support for real-time communication, persistent state, scheduled tasks, and tool integrations. Instead of cobbling together various services, agents-sdk delivers a cohesive platform where your agent code can immediately go into production on Cloudflare's global network.

Key Highlights:

1. Built-in state management - Agents come with integrated state management, including automatic synchronization between agent and clients, state change triggers, and direct SQL database access for each agent. No need to build separate state management systems or configure external databases.

2. Tool ecosystem - The framework includes ready-to-use tools for browsing the web, querying databases, calling AI models from any provider, and supporting human-in-the-loop approvals—all directly accessible through your agent code without additional configuration.

3. Workflow integration - Agents can trigger and coordinate asynchronous workflows that run for minutes, hours, or days with guaranteed execution and automatic retries. This enables complex sequences like file processing, database updates, and multi-step notification systems without building custom orchestration.

Quick Bites

Inception, an AI startup founded by a Stanford professor, has released Mercury Coder—the first commercial diffusion-based language model (dLLM) that generates text 5-10x faster than current LLMs. Traditional autoregressive models generate text sequentially, whereas Mercury's parallel "coarse-to-fine" approach achieves over 1000 tokens/ second on standard hardware. Mercury Coder matches or exceeds GPT-4o-mini and Gemini 2.0 Flash Lite on coding benchmarks. You can try Mercury Coder now in their new playground.

LangChain has released LangGraph Swarm, a lightweight Python library built on top of LangGraph, to create swarm-style multi-agent systems where specialized AI agents collaborate by dynamically handing off control based on their expertise. It maintains conversation context by remembering which agent was last active, and supports streaming, memory management, and human-in-the-loop capabilities. Installable with a simple pip install langgraph-swarm command.

Microsoft has expanded its Phi family of small language models (SLMs) with Phi-4-multimodal and Phi-4-mini. Phi-4-multimodal, a 5.6B parameter model, integrates speech, vision, and text processing in a powerful, compact, highly efficient model. Meanwhile, the 3.8B parameter Phi-4-mini is designed for text-based tasks, supporting long context, great at reasoning, math, coding, and function calling. Both models are now available on platforms like Azure AI Foundry, HuggingFace, and NVIDIA API Catalog.

OpenAI released GPT-4.5, their largest and best model for chat yet, in research preview, available to ChatGPT Pro users and via API, followed by Plus and Team users next week. The model represents a significant advancement in scaling unsupervised learning, resulting in improved natural interactions, broader knowledge, and better ability to follow user intent.

• GPT-4.5 focuses on scaling unsupervised learning rather than reasoning capabilities, which has led to stronger "EQ," reduced hallucinations, and better creative abilities for writing and design tasks.

• It is not a reasoning model — doesn’t think before it responds, which makes its strengths different from reasoning models. Compared to o1 and o3‑mini focused more on STEM, GPT‑4.5 is a more general-purpose, innately smarter model.

• GPT-4.5 API is extremely expensive: $75 for input and $150 for output per million tokens. To put things in perspective, the latest Claude 3.7 Sonnet API costs $3 for input (1/25th cost) and $15 for output tokens (1/10th cost).

• It supports web search, file uploads, and canvas for writing/code. It however currently lacks other multimodal capabilities such as Voice Mode, video, and screensharing — these will be integrated in future updates.

Tools of the Trade

1. Continue: Open-source platform to create, share, and use custom AI code assistants through IDE extensions. It offers a hub of pre-built "blocks" including models, prompts, rules, documentation, and tools that seamlessly integrate into existing development workflows.

2. YepCode: Run is a serverless runtime and SDK that executes AI-generated code in secure sandboxes without setup. It lets you run code securely, try packages without local installation, share runnable code examples, and test snippets quickly. It provides enterprise-grade security and scalability for executing arbitrary code.

3. Agenta: Open-source LLM Ops platform to build production-grade LLM applications faster and reliably. It provides end-to-end tools for the entire LLMOps workflow: building (LLM playground, evaluation), deploying (prompt and configuration management), and monitoring (LLM observability and tracing).

4. Awesome LLM Apps: Build awesome LLM apps with RAG, AI agents, and more to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos, and automate complex work.

Hot Takes

1. I can't stop thinking about the difference Deepseek R1 has already made
   if they continue to be transparent, open-source, and open-weights.....
   we will almost definitely get to UTOPIA and avoid the extremely dystopian outcome of AI being in the hands of a small number of oligarchs ~
   Bindu Reddy
   
   

2. We are in the era of $5 Uber rides anywhere across San Francisco but for LLMs weee ~
   Andrej Karpathy

That’s all for today! See you tomorrow with more such AI-filled content.

Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!

Unwind AI - X | LinkedIn | Threads | Facebook

Awesome LLM Apps | Sponsor Us

PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉