- unwind ai
- Posts
- Self-Developing AI Agent Framework
Self-Developing AI Agent Framework
PLUS: Claude 3.7 Sonnet and Claude Code agent, Perplexity's agentic web browser
Shubham Saboo & Gargi Gupta
February 25, 2025
Today’s top AI Highlights:
Fully automated and zero-code AI Agent framework
Opensource Plan Engine to prevent production failures in multi-agent apps
Claude 3.7 Sonnet hybrid reasoning model
Anthropic’s first agentic coding tool
Perplexity is releasing its own agentic-search web browser
& 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
Multi-agent apps are great but there are issues in the production environment, with agents crashing, message coordination failing, and expensive LLM calls. Orra is a coordination layer designed specifically for multi-agent AI applications that handles the complex orchestration between your AI agents and execution environment.
Orra acts as a "Plan Engine," sitting between your AI application and execution environment, handling complex interactions across various languages, frameworks, and platforms. The core design philosophy centers around ensuring production-grade reliability and preventing common pitfalls like cascading failures and excessive LLM costs experienced in dynamic agent workflows.
Key Highlights:
Framework-Agnostic - Works seamlessly with any agent framework in Python or JavaScript (with more languages coming soon), allowing you to keep your preferred development patterns while Orra handles the complex coordination underneath. You can build with your existing tools without forcing specific patterns or structures.
AI-Powered Planning - Uses AI to analyze intent and create execution plans that are validated against your domain constraints and capabilities. Plans are verified semantically before execution, preventing unsafe operations from running and ensuring all components work properly together.
Production-Ready Reliability - Implements exactly-once execution guarantees, automatic service health monitoring, and intelligent recovery mechanisms that prevent duplicate operations and provide clear audit trails. When services fail, Orra helps them reconnect and resume tasks automatically.
Tools as Services Architecture - Reduces latency and prevents hallucinations by running tools as persistent services that agents can access when needed, rather than loading them for each task. This approach also simplifies agent development by offloading complexity to specialized services.
AutoAgent is an open-source framework to build AI applications without writing code. It comes with a built-in team of AI agents - each specialized in tasks like coding, web browsing, and file management. You simply describe what you want to build in plain English, and these agents work together to create your AI application.
The framework has already proven its capabilities by ranking #1 among open-source solutions on the GAIA benchmark, delivering performance comparable to OpenAI's Deep Research agent. What makes it particularly useful is its flexible LLM support, working seamlessly with providers like Anthropic, OpenAI, Mistral, and Gemini.
Key Highlights:
Zero-Code Development - Build complete AI applications by describing what you want in natural language. The framework's agents handle everything - from writing code and creating tools to setting up workflows and integrating APIs. You can modify any part of your application later just by explaining the changes needed.
Ready-Made Agent Team - Comes with specialized AI agents that handle specific tasks: a Web Agent for browsing and search, a Code Agent for writing and executing code, a File Agent for managing documents, and an Orchestrator Agent that coordinates their work. Each agent has its own set of tools and operates in a secure environment.
Built-in Vector Database Management - AutoAgent comes with native vector database capabilities that automatically handle document processing and storage. The system takes care of converting various file formats, managing indexes, and optimizing retrieval - letting you focus on using the data rather than managing it.
Quick Start - Choose between two modes: User Mode for instantly using the pre-built multi-agent system, or Agent Editor Mode for creating custom agents and tools. Both modes work through natural language commands, making it accessible regardless of your technical expertise.
Production-Ready Architecture - Run locally during development and seamlessly deploy to containers or cloud environments. The system includes robust error handling, connection recovery, and audit logging out of the box. Importantly, all agent executions happen in secure sandboxed environments.
Quick Bites
Anthropic has released Claude 3.7 Sonnet, its most intelligent model to date and the first hybrid reasoning model on the market. While other companies have released reasoning models completely separate from their standard models, Claude 3.7 combines both Normal and Extended Thinking modes in one model.
Alongside the model, Anthropic introduced Claude Code, a command line tool for agentic coding, available as a limited research preview.
Claude 3.7 Sonnet:
Claude 3.7 Sonnet is available on all Claude plans and via API partners (Amazon Bedrock, Google Cloud Vertex AI) with the same pricing as previous models: $3/million input and $15/million output tokens. Extended thinking is available on all surfaces except the free tier.
You can pick when you want the model to answer normally and when you want it to think longer before answering. You can also see the model’s thinking process in this mode. In the standard mode, Claude 3.7 Sonnet is an upgraded version of the 3.5 Sonnet model.
API users can set specific token budgets for thinking, allowing for customized trade-offs between speed, cost, and answer quality.
The model is optimized lesser for math and computer science competition problems. It shows particularly strong improvements in coding, particularly front-end web development, and agentic workflows — outperforming OpenAI o1 and o3 models, and DeepSeek R1.
Claude Code:
This agentic coding tool in your terminal can search and read code, answer questions about architecture, edit files, write and run tests, commit and push code to GitHub, and use command line tools.
It understands your project context and takes real actions. No need to manually add files to context - Claude will explore your codebase as needed.
Planned improvements include enhanced tool call reliability, support for long-running commands, improved in-app rendering, and expanded understanding of its capabilities.
Perplexity has announced it's developing its own agentic-search web browser called "Comet," promising to "reinvent the browser" just as it did with search. While details about Comet's features and launch timeline remain under wraps, the company has opened a waitlist for the browser.
Google AI Studio now has a "Branch from here" feature for interacting with Gemini models. When iterating on prompts, you can select any message in a conversation, click "More Options," and create a separate, independent conversational branch to explore different directions without affecting the original thread. This branching allows for non-linear exploration, and you can easily switch between the original and branched conversations.
Tools of the Trade
agents.json: Open-source JSON specification built on OpenAPI that creates formal contracts for AI agent-API interactions, solving the problem of APIs being designed for developers rather than LLMs by introducing "flows" and "links" that describe how endpoints interact in sequence.
Mascotbot: Provides animated vector mascots for static AI agents and interfaces, that can serve as the visual interface for conversational AI systems. It provides both ready-made and custom-designed animated characters with real-time lip-sync capabilities that work with any TTS service.
Bhumi: High-performance Python client for LLM inference, offering 2-3x faster performance and 60% less memory usage than alternatives. Supports multiple providers including OpenAI, Anthropic, Gemini, and Groq. Also provides streaming capabilities, tool integration, and enterprise-ready features like rate limiting and error handling.
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
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