World Wire

Cloudflare has released the Agents SDK v0.5.0 to address the limitations of stateless serverless functions in AI development. In standard serverless architectures, every LLM call requires rebuilding the session context from scratch, which increases latency and token consumption. The Agents SDK’s latest version (Agents SDK v0.5.0) provides a vertically integrated execution layer where compute, state, and inference coexist at the network edge. The SDK allows developers to build agents that maintain state over long durations, moving beyond simple request-response cycles. This is achieved through 2 primary technologies: Durable Objects, which provide persistent state and identity, and Infire, a custom-built Rust inference engine designed to optimize edge resources. For devs, this architecture removes the need to manage external database connections or WebSocket servers for state synchronization. State Management via Durable Objects The Agents SDK relies on Durable Objects (DO) to provi...

Building AI agents is the new gold rush. But every developer knows the biggest bottleneck: getting the AI to actually communicate to your data. Today, travel giant Agoda is tackling this problem head-on. They have officially launched APIAgent , an open-source tool designed to turn any REST or GraphQL API into a Model Context Protocol (MCP) server with 0 code and 0 deployments . The Problem: The ‘Integration Tax ‘ Until recently, if you wanted your AI agent to check flight prices or look up a database, you had to write a custom tool. When Anthropic released the Model Context Protocol (MCP) , it created a standard way for Large Language Models (LLMs) to connect to external tools. However, even with MCP, the workflow is tedious. A developer must: Write a new MCP server in Python or TypeScript. Define every tool and its parameters manually. Deploy and maintain that server. Update the code every time the underlying API changes. Agoda team calls this the ‘integration tax.’ For a c...

In this tutorial, we build a human-in-the-loop travel booking agent that treats the user as a teammate rather than a passive observer. We design the system so the agent first reasons openly by drafting a structured travel plan, then deliberately pauses before taking any action. We expose this proposed plan in a live interface where we can inspect, edit, or reject it, and only after explicit approval do we allow the agent to execute tools. By combining LangGraph interrupts with a Streamlit frontend, we create a workflow that makes agent reasoning visible, controllable, and trustworthy instead of opaque and autonomous. Copy Code Copied Use a different Browser !pip -q install -U langgraph openai streamlit pydantic !npm -q install -g localtunnel import os, getpass, textwrap, json, uuid, time if not os.environ.get("OPENAI_API_KEY"): os.environ["OPENAI_API_KEY"] = getpass.getpass("OPENAI_API_KEY (hidden input): ") os.environ.setdefault("OPEN...

Alibaba Cloud just updated the open-source landscape. Today, the Qwen team released Qwen3.5 , the newest generation of their large language model (LLM) family. The most powerful version is Qwen3.5-397B-A17B . This model is a sparse Mixture-of-Experts (MoE) system. It combines massive reasoning power with high efficiency. Qwen3.5 is a native vision-language model. It is designed specifically for AI agents. It can see, code, and reason across 201 languages. https://ift.tt/yZtoL8k The Core Architecture: 397B Total, 17B Active The technical specifications of Qwen3.5-397B-A17B are impressive. The model contains 397B total parameters. However, it uses a sparse MoE design. This means it only activates 17B parameters during any single forward pass. This 17B activation count is the most important number for devs. It allows the model to provide the intelligence of a 400B model. But it runs with the speed of a much smaller model. The Qwen team reports a 8.6x to 19.0x increase in ...

The AI industry is currently obsessed with ‘agents’—autonomous programs that do more than just chat. However, most current multi-agent systems rely on brittle, hard-coded heuristics that fail when the environment changes. Google DeepMind researchers have proposed a new solution. The research team argued that for the ‘agentic web’ to scale, agents must move beyond simple task-splitting and adopt human-like organizational principles such as authority, responsibility, and accountability. Defining ‘Intelligent’ Delegation In standard software, a subroutine is just ‘outsourced’. Intelligent delegation is different. It is a sequence of decisions where a delegator transfers authority and responsibility to a delegatee. This process involves risk assessment, capability matching, and establishing trust. The 5 Pillars of the Framework To build this, the research team identified 5 core requirements mapped to specific technical protocols: Framework Pillar Technical Implementation Core...

In this tutorial, we build a fully stateful personal tutor agent that moves beyond short-lived chat interactions and learns continuously over time. We design the system to persist user preferences, track weak learning areas, and selectively recall only relevant past context when responding. By combining durable storage, semantic retrieval, and adaptive prompting, we demonstrate how an agent can behave more like a long-term tutor than a stateless chatbot. Also, we focus on keeping the agent self-managed, context-aware, and able to improve its guidance without requiring the user to repeat information. Copy Code Copied Use a different Browser !pip -q install "langchain>=0.2.12" "langchain-openai>=0.1.20" "sentence-transformers>=3.0.1" "faiss-cpu>=1.8.0.post1" "pydantic>=2.7.0" import os, json, sqlite3, uuid from datetime import datetime, timezone from typing import List, Dict, Any import numpy as np import faiss...