- unwind ai
- Posts
- Replit Agent using Claude 3.7 Sonnet
Replit Agent using Claude 3.7 Sonnet
PLUS: AI agent creates custom tools on the fly, Create datasets from AI search engine
Shubham Saboo & Gargi Gupta
February 27, 2025
In partnership with
Today’s top AI Highlights:
Personal, dynamic, and open-source AI agent that can create custom tools on the fly to complete any task
Pair on-device small models with LLMs on cloud — cut 80% API costs
Replit Agent V2 powered by Claude 3.7 Sonnet
Embeddings search agent with more than 20x recall than Google
Agent Command Centre to understand, plan, code, review and document software
& 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.
Latest Developments
Agent Zero is an open-source and fully customizable AI agent framework designed to grow and adapt alongside its user. It automates computer tasks by directly interacting with your operating system. Unlike AI chatbots that only provide text responses, Agent Zero can write and execute code, use terminal commands, browse the web, and create custom tools to solve problems.
It functions as a personal assistant that learns from each interaction, storing successful solutions in its memory to handle similar tasks more efficiently in the future. With its ability to break complex tasks into smaller components and delegate them to subordinate agents, Agent Zero can tackle everything from data analysis and content creation to system administration and development projects.
Key Highlights:
Computer as a Flexible Tool - Agent Zero relies on your OS rather than pre-programmed single-purpose tools. It writes its own code and uses the terminal to create custom tools for different tasks. The framework includes essential default tools like online search, memory features, and code execution, while letting you extend functionality with your own custom tools.
Multi-Agent Cooperation - It implements a hierarchical agent system where each agent reports to a superior (with humans at the top of the chain). Agents can create subordinate agents to handle subtasks, keeping context focused and improving complex problem-solving.
Fully Transparent Customization - Everything in the framework can be modified through accessible files - from the core system prompt to every message template and tool implementation. This transparency gives you complete control over agent behavior.
Persistent Memory and Learning - The system builds knowledge through continual use, storing previous solutions, code samples, and important facts to solve similar problems more efficiently in the future. It implements an advanced message history system that dynamically summarizes past interactions to maintain context while optimizing token usage.
Fully Dockerized - Agent Zero runs entirely within a Docker container, providing a consistent, secure, and isolated environment. You can deploy it anywhere Docker runs, and it can manage its own dependencies, including installing software and running code without impacting your host system.
The future of presentations, powered by AI
Gamma is a modern alternative to slides, powered by AI. Create beautiful and engaging presentations in minutes. Try it free today.
Processing large tasks with frontier models in the cloud can be prohibitively expensive—a single query on a million-token code repository costs over $15 with OpenAI's o1 model. Meanwhile, capable small LMs can now run on consumer devices, yet they're mostly used for simple tasks like text completion rather than data-intensive reasoning.
Together AI's Minions protocol addresses this gap by pairing small on-device models and frontier cloud models. Small language models running locally (via Ollama) handle the data-intensive parts of tasks like financial analysis and document processing, while frontier models provide advanced reasoning capabilities. This can reduce the costs by over 80% while maintaining 98% of the accuracy of cloud-only solutions.
Key Highlights:
Cost-Efficiency - Minions delivers 98% of frontier model performance at less than 18% of the cost. You can also opt for the simpler Minion protocol that offers 87% accuracy at just 3.3% of cloud costs, providing flexible options based on accuracy needs.
Task Decomposition - The cloud model generates code to break down complex tasks into smaller subtasks, which are then executed in parallel by the local model, maximizing hardware utilization. This addresses the difficulties small models face with long contexts and multi-step instructions.
Configurable and Customizable - You have control over key parameters, such as the number of communication rounds and the granularity of task decomposition.
Get Started - Works with Ollama for CPU devices or Tokasaurus for NVIDIA GPUs, with support for OpenAI and Together AI. The GitHub repo includes example code, Streamlit demo, and a notebook to help you integrate the protocol into your applications quickly.
Quick Bites
Replit is releasing Replit Agent v2, a more autonomous AI coding agent powered by Claude 3.7 Sonnet, featuring an industry-first realtime app design preview that renders interfaces as they're being built.
This new version forms hypothesis, intelligently searches files, and makes changes only when it has sufficient information, drastically reducing the likelihood of getting stuck in debugging loops. Currently available in early access for Replit paid plan subscribers who opt into Explorer Mode.
Here’s another banger from China. Stepchat has released Step-Audio, a groundbreaking open-source model family for intelligent speech interaction that integrates comprehension and generation capabilities in a 130B-parameter architecture. It supports multilingual conversations, emotional tones, regional dialects, adjustable speech rates, and prosodic styles including singing and rap. Licensed under Apache 2.0, completely free to use.
Exa AI has launched Websets, a search engine that deploys "an army of AI agents" to find 20x more accurate results than Google for complex queries. Unlike traditional search or even newer AI research tools like Deep Research, Websets excels at collecting comprehensive datasets where multiple criteria must be met - such as "Bay Area engineers with Rust experience" or "healthcare companies with technical founders." Keyword-based searches fundamentally cannot handle these complex, SQL-like queries that require extensive data collection across many websites.
Qwen just released QwQ-Max-Preview, a powerful new reasoning model built upon Qwen2.5-Max. This preview showcases impressive capabilities in math, coding, and agent-related workflows. The company plans to fully open-source both QwQ-Max and Qwen2.5-Max under Apache 2.0 license soon, alongside smaller variants like QwQ-32B for local deployment and native mobile apps for Android and iOS.
LMArena.ai has released Prompt-to-leaderboard (P2L), a custom leaderboard generated in real-time specifically for your prompt. Traditional leaderboards average performance across all tasks, P2L instantly tells you which model excels at your particular request - whether it's coding, math problems, or creative writing. Using P2L is simple - just enter your prompt in the new Prompt-to-Leaderboard tab on Chatbot Arena, and it immediately generates a leaderboard showing which models will handle your specific task best.
Tools of the Trade
MyCoder.ai: Open-source alternative to Claude Code. This command-line AI agent uses Anthropic's Claude API to perform tasks like creating projects, migrating code, and refactoring. It can parallelly execute tasks through sub-agents and self-modification.
ForeverVM: Code execution API to securely run arbitrary Python code in a remote sandbox and get back results. Unlike other code interpreters in which sessions state expires, ForeverVM uses memory snapshots to swap idle machines to disk for as long as you want. Applications and agents built on ForeverVM don't have to manage session lifecycles.
BoltAI: MacOS application that provides a unified interface to interact with multiple AI models like OpenAI, Anthropic, and Mistral, directly integrating AI assistance into various workflows. It offers customizable AI assistants, prompt libraries, and secure API key management.
Factory: AI-powered command center for software development, integrating code, documentation, and communication in a central hub. It uses AI agents to automate tasks, standardize workflows, and provides a unified context across the entire development lifecycle.
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
i’ve said this in private convos so I’ll say it here…
“better memory” is the final unlock we need to get truly better agents, and 2025 is when we’ll see more of this
we have strong reasoning, tools for tools, plenty of frameworks, but memory/context management needs improvement: grabbing the right context from long-term memory, summarizing memory to do this efficiently (i.e. graphs), privacy control (consumer) and user permissions (enterprise) as it relates to confidential memory, self-improvement (leveraging past memory to improve itself), etc. ~
Yohei NakajimaDeepResearch (the good one, OAI's).
DeepResearch (the OG one, Gemini's).
DeepResearch (the shit one, Perplexity's).
DeepSearch (Grok version -- equivalent to Deepseek's Search + DeepThink modes in nature). Substantially better than Perplexity's Pro Search.
o3m search (a top-end search function). Great, but outputs are very short and high-level.
Are the product people at ML SOTA labs just dumb or are they just unable to name things? ~
ludwig
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!
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 😉
Reply