The ReAct Loop: How Cursor & Claude Code Actually Work

0:05Welcome back to my course on AI agents from scratch In my last video I talked about tools and how a tool is how you transform a boring

0:17language model into an actionable ready AI agent Tools are things like program instructions JSON objects in hitting API calls

0:30CLI commands skills or just anything that allows the AI to actually execute actions or just get more information about a subject ChatGPT has access to

0:44tools but is still a chatbot it's still a back and forth conversation with a user To truly make an AI agent you have to make your agent at least

0:59somewhat autonomous In this video we're going to be talking about orchestration and how the orchestration layer is how you make your AI agent smart enough

1:13to actually do some sort of things It's honestly way easier than it sounds Now let's take this example again I updated the system prompt It can now

1:28do a couple of things Number one is it can generate strategies Number two is it can generate backtests And number three it can interpret backtest results Now let's pretend there's a little loop

1:42that automatically sort of executes this So we have this buy and hold spy strategy the AI responds with this strategy JSON Then we have this backtest command the AI responds with this different

1:56JSON which again we're going to send to an API We're going to get these performance results and we're going to feed it back into the AI agent Now the AI is summarizing the results

2:10and giving us an answer Imagine this loop done sort of by itself That is what makes an AI agent agentic

2:24And that's the key part of orchestration It is more or less literally just a loop a loop that has this format First the AI thinks about

2:38its current environment what is already done These observations are most of the time just the conversation context and other relevant details

2:51It will generate a thought and then we can generate an action We append the action we execute the action and we get the final result like the results we saw when I

3:05executed this down here And then we append it to the previous observations And then we repeat until we're done Done can mean many things It can mean a maximum number of iterations or

3:19a timeout or a specific tool that says hey we're done But the key idea is we'll repeat this loop until we've accomplished the user's goal

3:35This loop by the way it has a name It's called REACT and it's based on this paper Synergizing Reasoning and Acting in Language

3:48Models I recommend you take a look at this paper You don't have to read the entire thing but drop it in the chat GPT or Claude Read the abstract and understand because

4:01this is the basis of all AI agents today It's called React because the AI is literally reasoning it's thinking and it's acting

4:18And these things are simply LLM calls This is a system prompt designed to tell the LLM how to think This is another system prompt and they can

4:31be the same or separate It's just a system prompt that says given these thoughts what's the next best action This is a tool call You're executing the tool call whether it's a SQL query

4:46an API CLI tool whatever it is You're appending this results to the observations And then you're repeating over and over again All AI tools cursor Claude code

5:01all these AI tools are based on this framework And honestly this is it for a basic AI agent

5:17This is the key insight It is simply just structuring your LLMs to think and take actions Now this is an MVP It won't pass you the AI native software engineering

5:31interview because there's still so much more that we have to go through Like how do you trigger your AI How does your AI remember past events

5:44How does your AI improve How do you evaluate whether the AI is doing what you want it to do But all of these things are a little bit more advanced This here is the

5:57core of building AI agents and you don't even have to build it from scratch because there are libraries like LangChain that implements a lot of this for you You can just read the documentation

6:11and build your own AI agent In this next video we're going to be talking about how to make your AI agent production ready You can't just

6:24have this react loop and trust your AI to execute any action on its behalf Some tools might be expensive as in computationally expensive or monetarily

6:37expensive Some tools might just be slow Some tools might be dangerous such as the ability to edit files or delete files So you need some way of making your AI agent a little

6:51bit smarter and having an approval mechanism so sometimes it does things by itself and sometimes it asks for approval or help