The tutorial is comprehensive, covering the basic concepts and development history of AI agents, as well as practical projects, such as creating your own AI agent framework and building a multi-agent system. The course teaches practical skills such as memory management, contextual processing, communication protocols, and training methods for large language models. By following this tutorial, you can gain a deeper understanding of relevant knowledge while gaining real-world coding experience, from simply using AI models to designing intelligent systems independently. This will help you develop advanced AI skills that can be useful for job searching, research, or creating innovative AI applications. Relevant learning materials are published online and readily available at any time.
Early exposure to Agent.
AutoGPT、Multi-Agent、Planner、Tool Calling……
These concepts keep popping up on the timeline, but they’re more like a bunch of juxtaposed new terms than a logic that can be digested.
Everyone is doing “more automatic AI”,
But I don’t know:
How exactly does automaticism happen?
Until I read it again hello-agents.
At that moment, I realized: the model never changed
There is a very subtle, but very critical turning point.
Suddenly realized something:
Agents don’t introduce any new “sources of intelligence”.
The model is still the same model.
The capability boundary is not expanded.
There is also no mysterious increase in knowledge.
What really changes is the way the model participates.
It is no longer that:
Asked → give an answer
Instead, it is allowed to enter into a process that continues.
At its core, the Agent is actually a structure that allows you to think repeatedly.
Later I became more and more sure:
The core thing of Agent is actually very simple.
Here’s it:
The model is allowed to intervene in decisions multiple times rather than giving conclusions all at once.
This is completely different from chat mode.
Chat mode is more like:
“Tell me now what you think should be done at once.”
Agents are more like:
“Let’s take a step first, see the results, and then decide on the next step.”
When the model can:
- See what you did in the last step
- See what feedback the environment gives
- Then continue to judge based on this information
It’s no longer just a “language generator”.
Once that cycle appears, a lot of things go smoothly
For the first time, I focused on the loop itself, rather than what the model outputs. hello-agents
Think → Act → Observe → Think again
Once you acknowledge that the loop is the main body of the system and that the model is just a role in the loop, a lot of the previous twists are unraveled.
For example:
- Why do agents need state?
- Why can’t a prompt replace a process, no matter how well written?
- Why is “automation” essentially a transfer of decision-making power?
The answer is in that cycle.
The tool is actually secondary, the key is “who is choosing”
I was also attracted to Tool Calling before.
It seems that as long as there are enough tools, the agent is strong enough.
But then I slowly realized a more important question:
Who is deciding whether to use tools or not?
If it is a rule written by a person, it is just a script.
If the model makes judgments based on the current situation, it is the Agent.
The tools are just extensions.
The right to choose is the essence.
This is where I started to really differentiate between “automated processes” and “agent systems.”
Why this project is reassuring
hello-agents The feeling is not “amazing”, but reassuring.
It is not in a hurry to show the boundaries of competence,
Instead, it repeatedly shows you one thing:
The agent is not mysterious, it is just a program structure designed to allow the model to intervene repeatedly.
This made me feel for the first time:
Agent is not a trendy noun, but something that can be understood, disassembled, and reconstructed.
Looking at Agent again, it is a different perspective
Since then, when I look at any Agent framework, my attention unconsciously shifts to a few questions:
- How many times does the model appear in the system?
- Is it making judgments, or is it just filling in the blanks?
- Is the cycle open or artificially truncated?
- Are failures and deviations included in the process?
Instead of:
- How many tools are supported
- How complex a demo can run
GitHub：https://github.com/datawhalechina/hello-agents
Tubing: