Blog

Multi-agent systems are not just about having multiple agents. The real challenge is orchestration — how do agents pass data? How do they execute in parallel? How do you branch based on intermediate results? This article implements three orchestration patterns — sequential, concurrent, and conditional — using both LangGraph and qwen-agent, breaking down the core primitives and use cases of each.

Complex tasks can't be solved in one shot. The Planning Design Pattern separates 'thinking' from 'doing' — the Planning Agent decomposes user requests into a structured list of subtasks, and the Concierge Agent executes them in dependency order. This post dissects task decomposition, Pydantic structured output, and the planning-execution separation through a complete 7-day Paris trip case study.

Hand-writing system prompts is time-consuming, inconsistent, and hard to scale. The System Message Framework uses a single meta-prompt to let an LLM automatically generate system prompts — you only need to provide the role, company, and responsibility. This post dissects the two-step generation flow and explores its connection to building trustworthy AI Agents.

Traditional RAG follows a fixed retrieve-then-generate pipeline. Agentic RAG returns retrieval authority to the Agent — letting the LLM autonomously decide when to retrieve, what keywords to use, and whether a second retrieval is needed for verification. This post dissects Agentic RAG's core mechanism and the Producer-Checker iterative pattern through a complete travel knowledge base search case study.

When an Agent has multiple tools, how does the LLM orchestrate call order? How can downstream code reliably consume Agent output? How do you add human approval for side-effect operations (booking, charging)? This post covers three tool-use design patterns through a complete travel booking agent case study.

Building AI Agents isn't just about code — it's about design. This post covers three fundamental design patterns: using clear instructions to constrain Agent behavior, using Pydantic to make outputs verifiable, and splitting complex tasks with single-responsibility agents. Each pattern comes with runnable code comparisons.

Last time we hand-wrote a tool calling loop. Now we go one level up: what does an Agent framework actually do? We break it into four layers — Client, Agent, Tools, Session — mapping each from raw code to qwen-agent. You'll see that a framework is nothing more than engineered packaging of those 30 lines.

Strip away all frameworks. Using only the openai SDK connected to the Qwen API, hand-write the tool calling loop. In 30 lines of core code, understand: what an AI Agent really is, what Agent frameworks actually do under the hood, and why understanding this loop matters more than learning any framework.

Sharing how I built a modern personal homepage for AI researchers and developers using Next.js 16. This article covers technology selection, feature design, and how to leverage AI-assisted development.