Agent<\/a>And there's probably a real problem: it's not nothing, it's too much\u3002<\/strong><\/p>\nClaude Code, Codex, OpenCode, Gemini CLI, MCP, Skills, RAG, Memory, Multi-Agent, Computer Use, Browners Use... just figured out MCP today, and see Skills tomorrow; Claude Code, just learned you should study Multi-Agent\u3002<\/p>\n
Finally, there were hundreds of GitHub projects, but I never knew which one I should learn first\u3002<\/p>\n
Recently GitHub has an open source project worth collecting:i'm not sure if i'm gonna be able to do this<\/strong>.<\/p>\nAnd what it does is it's very simple, and it's just, you know, it's kind of like, it's kind of like, \"Ai Agent,\" and it's kind of kind of like, \"Ai Agent,\" and it's kind of like, you know, it's kind of like, \"Ai Agent,\" and it's kind of kind of like, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you know, you The project has been sorted out\u00a0240+AI\/Agent, 77 MCP\/ Skill, 23 basic hands-on exercises<\/strong>The entire learning system is divided into\u00a08 main phases, 2 core learning lines and follow-up to different professional directions<\/strong>.<\/p>\nMore importantly, it does not require all people to follow the same path, but first lets you understand a problem:<\/p>\n
Do you want to use Agent, or do you want to develop it<\/strong><\/p>\nThis question basically determines what you should learn in the next few months\u3002<\/p>\n
First, let's find out: What is AI Agent<\/strong><\/p>\nFor the first time many people have learned Agent, the biggest problem is to start putting frameworks in place. It's not really necessary. The first step is to figure out the basic logic of Agent's work\u3002<\/p>\n
Traditional ChatGPT works in a simple way:<\/p>\n
YOU ASKED FOR AN ANSWER\u3002<\/strong><\/p>\nFOR EXAMPLE, YOU ASKED ME TO ANALYZE THE VALUE OF BRITAIN'S RECENT INVESTMENTS. AI ANSWERS BASED ON CURRENT CONTEXT, AVAILABLE KNOWLEDGE AND AVAILABLE INFORMATION\u3002<\/p>\n
Agent goes further. When faced with a task, it does not simply generate an answer, but can accomplish several steps in a continuum around the goal:<\/p>\n
Understanding the mission poaching task poaching the call tool poaching data poaching the results of the analysis poaching the next judgement continue the final delivery\u3002<\/strong><\/p>\nFor example, you told one Agent that you would help me study the fundamental changes that have taken place in Britain in the last quarter and put together an investment study\u3002<\/p>\n
An Agent, with the appropriate tools and data access capabilities, can continue to search for company information, obtain financial information, read financial statements, extract income and profit data, find relevant news, analyse industry changes, compare competitions and eventually compile this information into studies\u3002<\/p>\n
So understand Agent, the most important thing:<\/p>\n
Agent's value is not just AI more chatable, but AI is beginning to be able to call tools and perform successive tasks\u3002<\/strong><\/p>\nWhen you figure this out, it's easy to see the Tool Use, MCP, Skills, Memoory and Multi-Agent in the back\u3002<\/p>\n
IAgent, don't come up and study Multi-Agent<\/strong><\/p>\nThis is one of the most valuable places on this learning map. It doesn't make you learn complex frameworks at once, but it's tearing the Agent learning process apart\u3002<\/p>\n
The overall learning order can be understood as:<\/p>\n
LLM & Prompt & Tool Use & Agent MCP \/ Skills & Memory \/ RAG & Multi-Agent & Company Use \/ Brown Use \/ Sandbox\u3002<\/strong>\u00a0<\/strong><\/p>\nThe easiest mistake here is jumping\u3002<\/p>\n
Python, for example, didn't know what the API was, and JSON couldn't read it, so he started to study AutoGen, Crewai, LangGraph and Multi-Agent. It must have been painful to learn and it was easy to give up in a few days\u3002<\/p>\n
So before the official start, it is important to determine which of us belongs\u3002<\/p>\n
III. Route A: CLI Power User<\/strong><\/p>\nThe first is actually a very large category of people\u3002<\/p>\n
You don't want to be Agent Engineer, and you don't have to develop an Agent from zero. What you really want to do is:It\u2019s not easy to get Claude Code, Codex, OpenCode, Gemini CLI, who are now Agent\u3002<\/strong><\/p>\nFor example, writing codes, making websites, writing articles, analysing information, processing Excel, doing research, auto-managing documents, calling browsers, connecting GitHub, processing its own knowledge base, or handing over daily duplicated work to AI\u3002<\/p>\n
If that's what you need, you don't have to spend a few months studying Agent Framework\u00a0Track A: CLI Power User<\/strong>.<\/p>\nThe official study time on this route is about..\u00a08-10 weeks<\/strong>.<\/p>\nIV. Phase I: Completing the foundations<\/strong><\/p>\nThis part of the road map\u00a0Stage 0: Foundations<\/strong>It's suggested to leave one to two weeks\u3002<\/p>\nYou don't have to train yourself as a professional programmer for learning Agent, but the basic concepts of Terminal, Git, API, JSON are best understood, otherwise MCP, Agent, or GitHub are often stuck on very basic issues\u3002<\/p>\n
Terminal \/ CLI<\/p>\n
at least know what the terminal is, what the cd, ls, mkdir, git cline are probably doing\u3002<\/p>\n
No back orders, much less programmers. The real goal is just:It's not entirely strange to see someone else turn you on and enter an order\u3002<\/strong><\/p>\n2. Git and GitHub<\/p>\n
At the very least, there is a need to understand the most common concepts: Repository, Clane, Commit, Push, Pull, Branch\u3002<\/p>\n
The simple reason is that a large number of Agent, MCP Server, Skills and Open Source tools are now in GitHub. In the future, you install a Skill, deploy a MCP Server, or study an open source Agent, which is almost impossible to bypass GitHub\u3002<\/p>\n
3. API<\/p>\n
API'S A VERY IMPORTANT TURN\u3002<\/p>\n
At least we should understand what API Key, Endpoint, Request, Response, Token and JSON mean\u3002<\/p>\n
Claude Code Mode Model, Agent Mode Search Service, MCP External System, in essence, has a large number of operations. If you don't even know what API Key is, it'll be hard to learn later\u3002<\/p>\n
4. JSON<\/p>\n
JSON DOESN'T NEED TO STUDY FOR A LONG TIME. ALL YOU NEED TO SEE IS THE FOLLOWING STUFF, AND YOU PROBABLY KNOW IT'S STRUCTURED DATA:<\/p>\n
{
\n\"name\": \"Kelvin\",
\n\"task\": \"research\",
\n\"model\": \"claude\"
\n}<\/p>\n
Later on, when you configure MCP, Agent, API, you often encounter similar content\u3002<\/p>\n
V. PHASE 2: UNDERSTANDING HOW LLM WORKS<\/strong><\/p>\nThis part corresponds\u00a0Stage 1: LLM Basics<\/strong>It is enough that ordinary users spend about a week understanding the basic concept\u3002<\/p>\nThere is no need at this stage to study complex mathematical formulas, focusing on the relationship between Token, Context Windows, models, and Agent\u3002<\/p>\n
Token can be easily understood as the basic unit used to process the text of the model, which affects the length of the context, the API costs and the way parts of the model are called\u3002<\/p>\n
The following is a simple translation of the text:AI HOW MUCH CONTEXT INFORMATION CAN BE PROCESSED AT A TIME\u3002<\/strong>\u00a0The concept continues to be felt in the back of Memoory, RAG, Context Engineering\u3002<\/p>\nAnother very important difference is not to mix models with Agent\u3002<\/p>\n
This can be roughly understood as:<\/p>\n
LLM is brain, while Agent is brain + tool + workflow + execution environment\u3002<\/strong><\/p>\nClaude, GPT, Gemini is a model system, while Claude Code, Codex, a product like this, further places the model in the Agent workflow that can operate tools, codes and the environment\u3002<\/p>\n
VI. Phase 3: Prompt Engineering<\/strong><\/p>\nNow it's often said Prompt is out of date, not really. After entering the Age, Prompt's role changed\u3002<\/p>\n
You've studied Prompt in the past, probably looking at how to write a very powerful hint. More importantly now:How to give Agent a clear mandate, sufficient context, clear rules and final acceptance standards\u3002<\/strong><\/p>\nA more practical Agent command should consider at least four parts: mission, Context, Constraints and Outlook\u3002<\/p>\n
Like you can tell Agent:<\/p>\n
\n\n- Study the operations of a company in the last quarter. Priority is given to the use of corporate financials and official announcements, all key figures need to be sourced, and an investment study in the Markdown format is eventually generated and potential risks are identified separately\u3002<\/li>\n<\/ul>\n<\/blockquote>\n
THIS WRITING IS NO LONGER JUST A SIMPLE QUESTION FOR AI, BUT IS DEFINING A FULL MISSION\u3002<\/p>\n
VII. Formal access route A: Select a CLI Agent<\/strong><\/p>\nOnce the basic learning ahead is completed, the real use of Agent can begin\u3002<\/p>\n
The easiest mistake here is Claude Code, Codex, Gemini CLI, OpenCode, all installed, and which one is better every day\u3002<\/p>\n
Not really\u3002<\/p>\n
Pick one and use it for a week\u3002<\/strong><\/p>\nLike choosing Claude Code, or choosing Codex. Don't rush into complex MCP and Multi-Agent this week, but let it do the real thing\u3002<\/p>\n
You can make it help you create a website, read the information in a folder, analyze the structure of a GitHub project, modify the program, organize research materials, or generate a Markdown report based on a collection of materials\u3002<\/p>\n
The core objective of this phase is one:<\/p>\n
It's customary to give a full mission to Agent, not just to use it as a chat robot\u3002<\/strong><\/p>\nVIII. Create your own Agent workflow<\/strong><\/p>\nWhen you can use Claude Code or Codex properly, you start to have a really valuable phase:Workflow\u3002<\/strong><\/p>\nThe previous way to use AI may be to open ChatGPT, ask a question, copy the answer, paste to another place and continue\u3002<\/p>\n
Agent's modus operandi should gradually become: you give it a goal, read the entire project and context, develop an implementation plan, then start operations, check and modify when implementation is completed, and deliver the results\u3002<\/p>\n
This is a time to start learning about CLAUDE.md, AGENTS.md, project rules, reusable Prompt, task dismantling and project context management\u3002<\/p>\n
For example, you can create your own content area by writing a lot:<\/p>\n
contact-workspace\/
\nresearch\/
\n- draft\/
\nideas\/
\nideas - publiced\/
\nCLUDE.md<\/p>\n
And then, in the CLUDE.md:<\/p>\n
\n\n- All articles are available in Chinese. Check the facts and start writing. Technical content gives priority to official documents. The article avoids the marketing cavity and the obvious AI taste, and when it is finished, the output of Marktown is unified\u3002<\/li>\n<\/ul>\n<\/blockquote>\n
After that, every time that Agent entered the project, he could work according to relatively fixed rules\u3002<\/p>\n
AND THAT'S WHEN YOU REALLY STARTED WITH A SIMPLE AI CONVERSATION\u00a0AI WORKFLOW<\/strong>.<\/p>\nIX. STARTING LEARNING MCP, Skills, Plugins and Subagents<\/p>\n
At Stage 5, the capacity of Agent will begin to change significantly\u3002<\/p>\n
This is also where many people are now the easiest to mix several concepts\u3002<\/p>\n
WHAT'S MCP<\/p>\n
MCP CAN BEGIN WITH A SIMPLE INTERPRETATION:<\/p>\n
To solve the problem of Agent's connection to external tools and data\u3002<\/strong><\/p>\nAgent is more capable of doing what it can do without access to your GitHub, Google Drive, databases, browsers or internal operating systems\u3002<\/p>\n
Through MCP, Agent can be allowed to establish standardized links with different tools and data sources\u3002<\/p>\n
SO THERE'S NO NEED TO LOOK AT THE BOTTOM OF THE MCP AGREEMENT IN THE FIRST PHASE\u3002<\/p>\n
You just have to figure out one thing:<\/p>\n
MCP deals mainly with Agent's connection to the outside world\u3002<\/strong><\/p>\nAND THEN YOU CONNECT YOURSELF TO ONE OR TWO OF THE SERVICES YOU REALLY NEED, WHICH IS MORE EFFECTIVE THAN WATCHING TEN MCP PRINCIPLES ARTICLES\u3002<\/p>\n
What is Skills<\/strong><\/p>\nIf MCP is more biased, Agent\u00a0What does it connect to<\/strong>Then Skill is closer to solving, Agent\u00a0How do you do something like that<\/strong>.<\/p>\nFor example, you can do a stock analysis, Skyll, which requires that Agent, at the time of analysis, do so according to a fixed process: read the financial statements, then analyse the income, profits and cash flows, then examine the valuation, the industry cycle and the potential risks, and then produce the report in a uniform format\u3002<\/p>\n
You don't have to rewrite dozens of lines from now on\u3002<\/p>\n
From this perspective, it is simply understood that:<\/p>\n
LLM provides basic intelligence, Agent is responsible for implementation, MCP provides external connectivity, and Skill provides professional working methods\u3002<\/strong><\/p>\nWhen these things are put together, a universal Agent begins to become a professional AI assistant in a certain field\u3002<\/p>\n
What's Subagent<\/strong><\/p>\nWhen the task is getting more complicated, it's not always the best way for a person to get to the bottom of it\u3002<\/p>\n
For example, by allowing an Agent to search for information, write codes, analyse data, and review results, the complexity of the task can easily lead to context confusion\u3002<\/p>\n
So you can untangle the task:<\/p>\n
Master Agent
\nResearch Agent
\nCoding Age
\nIdeas - Data Agent
\nRevieAgent<\/p>\n
The main Agent is responsible for overall coordination and the different Subagents are responsible for their own tasks\u3002<\/p>\n
This is also the basis for understanding Multi-Agent later\u3002<\/p>\n
XII. COMPUTER USE, Browser Use and Sandbox<\/strong><\/p>\nIt's a very important direction for Agent's development now\u3002<\/p>\n
While many AI systems used to call external services mainly through API, Agent has now started to acquire further implementation capabilities such as browsers, computers, terminals, code running environments and Sandbox\u3002<\/p>\n
For example, a full future research mission might be: Agent automatically opens the browser for information, downloads files, reads data, runs Python, generates charts, and finally organizes a report\u3002<\/p>\n
When Agent actually got into these implementation environments, it began to move from a talking robot that simply answered questions to a digital worker who could do the job\u3002<\/p>\n
Thirteen, if you just want to work more efficiently, it's enough to learn about it<\/strong><\/p>\nIT'S VERY IMPORTANT BECAUSE A LOT OF PEOPLE LEARN AI AND EVENTUALLY HAVE AN UNNECESSARY ANXIETY\u3002<\/p>\n
When you look at LangGraph, you think you should learn; when you look at Multi-Agent, you're afraid you're behind\u3002<\/p>\n
In fact, if your goal is mainly AI writing, AI programming, AI research, AI office, data analysis and automation, then the real use of Claude Code or Codex, MCP, Skills, Workflow and the underlying Computer Use can solve a lot of practical problems\u3002<\/p>\n
Not everyone needs to be Agent Engineer\u3002<\/strong><\/p>\nIf Agent is already able to solve your problems, use it to the extreme\u3002<\/p>\n
ONLY WHEN YOU ACTUALLY FIND THAT AN OFF-THE-SHELF TOOL DOESN'T MEET THE DEMAND WILL YOU ENTER ROUTE B\u3002<\/p>\n
XIV. Route B: Agent Builder<\/strong><\/p>\nIf your goal is not to use Agent, it isDevelop the Agent product or Agent system itself<\/strong>Then we need to keep going back to Track B\u3002<\/p>\nThis route will become significantly more technical, with the following broad learning order:<\/p>\n
@LLM_Prompt \u2192Tool Use \u2192AgentFramework \u2192MCP\/Skills \u2192Memory \/ RAG \u2192Multi-Agent \u2192Computer Use\u3002<\/strong>
\nTHE BASIC LLM QUOTE TOOL USES THE QUEST AGENT FRAMEWORK MCP\/ SKILLS MEMORY \/ RAG \u2192 MULTI-AGENT COMPUTER USE\u3002<\/strong><\/p>\nThe real watershed is Tool Use\u3002<\/p>\n
15, Tool Use: Agent Builder's real starting point<\/strong><\/p>\nYou need to start understanding why the model calls the tool and how it judges when it should be\u3002<\/p>\n
FOR EXAMPLE, YOU GAVE AI A FEW TOOLS:<\/p>\n
seech_web()
\nSearch Network ()
\nread_file()
\nRead Files
\nwriter_file()
\nrun_python()
\nRun Python function<\/p>\n
And tell it: help me study Tesla's recent financial situation\u3002<\/p>\n
The model will need to determine when to search for information, when to read documents, when to run Python, when to continue, and when the task has been completed\u3002<\/p>\n
This is the core of Tool Use\u3002<\/p>\n
When you really understand this layer, it's easier to study Agent Framework\u3002<\/p>\n
16, Agent Framework: Don't learn five at a time<\/strong><\/p>\nAfter entering Agent Framework, you will see a lot of frames like LangGraph, AutoGen, CrewAI, OpenAI Argentinas SDK, Pydanticai\u3002<\/p>\n
The easiest mistake here is to learn everything\u3002<\/p>\n
The right wayPick a frame and do a real project<\/strong>.<\/p>\nLike doing an AI stock study, Agent. Once a stock is entered, Agent automatically obtains company information, accesss financial data, analyses relevant news and industry changes, conducts valuation analyses and produces studies\u3002<\/p>\n
It would be more valuable to actually complete such a project than to look at the curriculum of five Agent Framework at the same time\u3002<\/p>\n
XVII. MEMORY AND RAG<\/strong><\/p>\nWhen Agent started working, two questions came up very quickly: how can he remember the past? How does it find what is really relevant from a lot of its own information<\/p>\n
The first question corresponds\u00a0Memory<\/strong>.<\/p>\nMemoory can simply be understood as allowing Agent to keep important information from the past, such as user preferences, historical tasks, long-term status or some key context\u3002<\/p>\n
The second question corresponds\u00a0RAG<\/strong>.<\/p>\nAssuming you have 1,000 PDFs, 500 articles, a lot of corporate financials, industry reports and minutes, it is not possible to put everything in Prompt every time you ask a question\u3002<\/p>\n
THE BASIC LOGIC OF RAG IS:<\/p>\n
The user asks questions \u2013 the system searchs the knowledge base \u2013 and finds the relevant information \u2013 and puts it in the Context \u2013 LLM\u3002<\/strong><\/p>\nSO THE REAL SOLUTION FOR RAG IS:HOW TO GET AI TO FIND THE INFORMATION THAT IS REALLY NEEDED FROM A LOT OF OUTSIDE KNOWLEDGE\u3002<\/strong><\/p>\n18, Multi-Agent: It makes sense to come here<\/strong><\/p>\nIt's natural to go back to Multi-Agent after the first Tool Use, Agent Framework, Memory and RAG\u3002<\/p>\n
For example, if you want an AI investment research team, you can design Research Agen for information, Financial Agen for financial analysis, Industrial Agen for industry research, Valuation Agen for valuation, Risk Agen for risk-seeking, and ultimately Child Investment Agen for all results\u3002<\/p>\n
The system will then evolve from an Agent to a collaborative AI team\u3002<\/p>\n
But special attention should be paid to:<\/p>\n
Multi-Agent is not as much as Agent\u3002<\/strong><\/p>\nIf an Agent can complete the task, there is no need for a Multi-Agent and Multi-Agent. A lot of Agent is of real value only if the task can really be broken down into several parts that are relatively independent, require different contexts or different professional abilities\u3002<\/p>\n
19. Which route should ordinary people choose<\/strong><\/p>\nIt's really simple to judge\u3002<\/p>\n
IF YOU'RE A PRODUCT MANAGER, AN OPERATOR, A SALESER, AN INVESTMENT RESEARCHER, A CONTENT CREATOR, AN ENTREPRENEUR, AN INDEPENDENT DEVELOPER, OR A COMMON AI USER, YOU SHOULD BE GIVEN PRIORITY\u00a0Track A<\/strong>.<\/p>\nFirst put:<\/p>\n
Claude Code \/ Codex + MCP + Skills + WorkFlow<\/strong><\/p>\nReally use it up\u3002<\/p>\n
If it is discovered that the ready-made Agent has failed to meet its needs, then go back to Track B and continue to study Tool Use, Agent Framework, Memory, RAG, Multi-Agent and Company Use\u3002<\/p>\n
It would be much more efficient to learn\u3002<\/p>\n
20 for ordinary people, 10 for practical learning<\/strong><\/p>\nIf I start learning again today, I'd rather do it at this pace, instead of chasing new tools every day\u3002<\/p>\n
Week 1: Foundation\u3002<\/strong>\u00a0Study Terminal, Git, Github, API and JSON. The goal is not to learn programming, but to be able to understand basic commands and run some GitHub open source projects\u3002<\/p>\nWEEK 2: LLM\u3002<\/strong>\u00a0Token, Context, API, Prompt, and the difference between models and Agent. The goal is to create a basic understanding that does not require the study of complex theories\u3002<\/p>\nWeek 3: CLI Agent\u3002<\/strong>\u00a0Claude Code and Codex choose one as the main tool to give it at least one real task every day, not just Demo\u3002<\/p>\nWeek 4: Workflow\u3002<\/strong>\u00a0Started learning about CLAUDE.md, AGENTS.md, project context, task dismantling and workflow, and created its first AI Workspace\u3002<\/p>\nWEEK 5: MCP\u3002<\/strong>\u00a0Do not install dozens of MCP Servers at a time, connect one to two data sources or tools that you really will use\u3002<\/p>\nWeek 6: Skills\u3002<\/strong>\u00a0A simple Skill, which can be an article study, a stock analysis, a sales analysis, a Skill, or any Skill related to your actual work\u3002<\/p>\nWeek 7: Subagent\u3002<\/strong>\u00a0Find a complex task, break it down into two or three characters, and experience the way Agent works with Subagent\u3002<\/p>\nWeek 8: Browner\/Computer Use\u3002<\/strong>\u00a0Start trying to get Agent to operate a browser, code environment, or Sandbox, and bring Agent from the pure text world into the real execution environment\u3002<\/p>\nWeek 9: Full project\u3002<\/strong>\u00a0THE STUDY OF THE NEW CONCEPT IS SUSPENDED AND ONLY A TRULY OPERATIONAL PROJECT, SUCH AS AI INVESTMENT RESEARCH DESK, AI CONTENT PRODUCTION DESK OR AI SALES ASSISTANT\u3002<\/p>\nWeek 10: Optimization\u3002<\/strong>\u00a0Instead of pursuing new tools, we look back at our whole system: which tasks can be automated? Which Prompt can become Skill? What data should be accessed through the MCP? What jobs can be given to Agent? What are the key elements that must be manually reviewed<\/p>\nIf these five questions are clear, you're starting to really understand Agent\u3002<\/p>\n
XXII. Learning Agent's most important method: learning to do<\/strong><\/p>\nDo not study Agent in the same way as you read the curriculum, keep the collection, and keep the collection\u3002<\/p>\n
The biggest problem with this learning approach is that it appears to be going on a daily basis and in fact has not changed in its own way of working\u3002<\/p>\n
More effective methods are:<\/p>\n
Learn a concept to do a thing to do a problem to learn a concept to solve problems with new knowledge\u3002<\/strong><\/p>\nLearn Prompt, write a real and repeated Prompt; learn Claude Code, finish a project with it; learn MCP, connect to a service you need; learn Skyll, write a Skyll; learn RAG, make a knowledge base; learn Multi-Agent, break a real mission into two Agents\u3002<\/p>\n
Don't think of learning as a goal\u3002<\/strong><\/p>\nTwenty-two, what does this AI Agent really solve<\/strong><\/p>\ni think the most valuable place that it collects is 240+\u3002<\/p>\n
GitHub has never been short of projects, and the AI era is not missing\u3002<\/p>\n
What is really scarce is:Learning order\u3002<\/strong><\/p>\nIt actually helped us answer three important questions: where am I now? What should we learn next? What level of learning to stop first<\/p>\n
In particular, the last question\u3002<\/p>\n
Not everyone needs to be Agent Engineer\u3002<\/p>\n
The ability to use AI Agent in the future can be roughly divided into three levels\u3002<\/p>\n
First floor is..Agent is used<\/strong>I don't know. Can really integrate Claude Code, Codex and Agent into their work\u3002<\/p>\nSecond floorCan configure Agent<\/strong>I don't know. It is possible to use MCP, Skills, Memory, WorldFlow to transform the UniversalAgent into an AI working system for itself\u3002<\/p>\nThird floorCould build Agent<\/strong>I don't know. Tool Use, Agent Framework, RAG, Multi-Agent, Computer Use, etc\u3002<\/p>\nFor the vast majority of ordinary people, it is possible to achieve very significant efficiency gains by first level and then second level\u3002<\/p>\n
We'll go to the third level when we really meet problems that are not solved by available tools\u3002<\/p>\n
It's a lot more realistic than getting all kinds of Agent Framework in the first place\u3002<\/p>\n
at last<\/p>\n
and if you open it for the first time, it's not recommended to go straight into the 240+ project and look at it\u3002<\/p>\n
Make sure one thing first:<\/p>\n
Do you want to use Agent, or do you want to develop Agent<\/strong><\/p>\nIf you just want Claude Code, Codex, MCP, Skills, go\u00a0Track A<\/strong>.<\/p>\nIf the ultimate goal is to develop the Agent product, then go\u00a0Track B<\/strong>.<\/p>\nIf we don't know which one we belong to now, it's simple:<\/p>\n
GO FIRST A\u3002<\/strong><\/p>\nLet Agent really finish you off\u3002<\/p>\n
When you start to meet the boundaries of the ready-made Agent, you will know why you need to learn Tool Use, RAG, Memoory and Multi-Agent\u3002<\/p>\n
AI Agent is really worth learning, not today, but:<\/p>\n
Can you put these things together and really solve problems for yourself\u3002<\/strong><\/p>","protected":false},"excerpt":{"rendered":"If you start studying AI Agent recently, there's probably a real problem: it's not that there's nothing to learn, it's that much. Claude Code, Codex, OpenCode, Gemini CLI, MCP, Skills, RAG, Memory, Multi-Agent, Computer Use, Browners Use... just figured out MCP today, and see Skills tomorrow; Claude Code, just learned you should study Multi-Agent. Finally, there were hundreds of GitHub projects, but I never knew which one I should learn first. Recently Githu<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[144],"tags":[1405,3765,8811],"collection":[],"class_list":["post-56059","post","type-post","status-publish","format-standard","hentry","category-baike","tag-agent","tag-ai","tag-8811"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/56059","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/comments?post=56059"}],"version-history":[{"count":0,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/56059\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/media?parent=56059"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/categories?post=56059"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/tags?post=56059"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/collection?post=56059"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
![\"ca170ffj00tk24o90ind000v90cip\"](\"https:\/\/www.1ai.net\/wp-content\/uploads\/2026\/08\/ca170fbfj00tk24o900ind000v900cip.jpg\" "\\"ca170ffj00tk24o90ind000v90cip\\"")
<\/p>\n