November 24, 2025 - Day 2: Agent Design, Frameworks & Life Reflections

Today I explored the second section of the Hello Agents open-source guide, focusing on Agent Design. Compared with yesterday's foundational theory, today's content was much more practical, involving real agent patterns, reasoning methods, design choices, and hands-on coding.

The deeper I went, the more I realized how differently each agent method behaves — not just in theory, but in real execution.

🧩 Core Agent Methods Learned Today

🔁 1. ReAct — For Uncertain, Interactive, Tool-Based Tasks

ReAct (Reasoning + Acting) excels when:

• Task goals are uncertain
• You need the agent to interact with external APIs or websites
• You want transparent reasoning + action steps

It's flexible and intuitive — a good default for tasks that require "thinking while doing."

🧭 2. Plan-and-Solve — For Clear Logic & Structured Breakdown

Best for tasks that emphasize internal reasoning and structured planning:

• 🧮 Math tasks
• 📊 Reports that require multi-source data
• 💻 Code generation & completion

It feels more stable because the agent commits to a plan before acting — a bit like writing an outline before writing an essay.

🔍 3. Reflection — For High-Quality, Reliable Results

Reflection adds a self-review loop. It's ideal when accuracy and reliability matter.

Example from the course:

编写一个 Python 函数找出 1 到 n 之间所有素数（prime numbers），并通过反思降低复杂度。

Reflection makes the agent refine its own output — almost like peer review, but done by itself.

🛠️ Hands-On Projects: What I Built Today

I tried running frameworks like Agentscope, AutoGen, and more. Some examples were surprisingly fun:

• 💰 A Bitcoin price checker using AutoGen
• 🎮 A mini Sanguosha (三国杀) agent using Agentscope

There are many other frameworks too — CAMEL, AgentVerse, LangChain — but today's examples helped solidify the agent loop and tool-calling intuition.

⚙️ Real-World Problems I Faced Today

While practicing, two major challenges became obvious:

1. 📦 Libraries evolve ridiculously fast

A lot of tutorial code breaks because:

• Function names change
• APIs deprecate
• Parameters get updated

You constantly need to adjust the scripts.

2. 🔑 Everything needs API keys

LLM calls, AMap, search engines, external data... Every. Single. Thing. needs a key.

Maintaining + organizing keys becomes surprisingly resource-intensive.

These two issues made me realize that agent development isn't just coding — it's environment engineering.

🧩 Slowing Down & Replanning

Day 2 had so much content that I honestly feel I need to revisit it again:

• Too much theory
• Too much code
• Too many frameworks
• Too many APIs

Tomorrow, I'll switch to low-code platforms (Coze, Dify, n8n) before returning to the more complex frameworks.

I was too optimistic trying to finish the entire course in 5 days. But learning should have breath.

💬 Final Thought of Day 2

It's okay to slow down. Understanding deeply matters more than finishing quickly.

Today was a comprehensive technical lesson covering agent design patterns, practical implementations, and real-world development challenges.

🎯 My Learning Progress

Progress Bar: ■■■■□□□□□□ (40%)