Agentic Coding with Claude Code | Agentic AI Coding Tutorial | Coding With Claude Code | Simplilearn

Key Takeaways

[Music] Hello everyone and welcome to today's tutorial on agentic AI coding with clot. In this tutorial, we'll be building an agentic coding assistant powered by cloud's anthropic AI. It's like creating your own AI developer that can read and write code files, search through codebase, run terminal commands like npm, get ls, etc., install dependencies and run tests and most importantly it is going to decide what to do next on its own. So, so we are going to build all of this with the help of NodeJS. Now, if you want to learn more about it, watch this video till the end. Now, before you move ahead, here's the agenda of our today's session. We are going to start with an introduction. Then, we are going to understand our folder structure of this project. Moving ahead, we are going to also create subfolders in the respective folder structure. And finally, we are going to run our project. Before we commence, if you are interested in mastering the future of technology, the professional certificate course in generative AI and machine learning is a perfect opportunity for you. This course is offered in collaboration with ENICT Academy and IIT Kpur. And this 11-month live and interactive program provides hands-on expertise in cuttingedge areas like generative AI, machine learning and tools such as chat GPTU and hugging face. You're going to gain practical experience through 15 plus projects, integrated labs and live master classes delivered by esteemed IIT Kpur faculty. Alongside earning a prestigious certificate from IIT Kpur, you'll also receive official Microsoft badges for Azure AI courses and career assist through simply learns job assist program. So guys, hurry up now and join the course. The course link is mentioned in the description box. Now before you move on, there's a short quest to test your knowledge. So guys, can you tell which team created Claude? Your options are OpenAI, Anthropic or Google or Microsoft? Please mention your answers in the comment section below. Now before we start the tutorial I request you guys that do not forget to hit the subscribe button and click the bell icon so that you don't miss out any update from our end. So let's get started. So guys let us first understand what is claude. Claude is a next generation AI platform which helps you in assistance. So you can type anything. For example, you can take a code from your given project. Let's say it is somewhere around a Python script and you are facing some errors. Just put it on cloud and the next thing you get is is the output. So you can see it's a very amazing tool. It helps you in a lot of task. So in this project what we will be doing is we are going to take the API key from the cloud. Okay. So if you want to try out cloud just right click all over here and uh if you want to search the API key so you can just type cloud API and uh just navigate to cloud developers platform okay so there you can get your own API for this project you'll be requiring an API key for this the chargeable fees would be around $5 so you can go to the claude's website and you can just generate your own API I key. So you can see all over here after right clicking you know it is going to generate your API key and the same API key we are going to use in our project. So I hope so guys it's clear to you. Now in this project what we are basically doing is you can see all over here we are going to create a assistant with the help of claude and you can see there is a one file sandbox which will be you know the files will be generated all over here by the chat assistant. So in this the cloud is going to act as an anonymous software engineer by thinking and planning and also calling tools you know to you know approach the given problem refine the approach also and you it is going to do the specific task as you assign it step by step. So you would have a question right now what exactly are we building? So we are bu setting up a local tool calling framework which includes run agent a loop that sends your task to claude like let's say for example you ask it to create a todo app and uh this agent is going to do it. It is also going to execute uh tools such as read write um you know shell commands and all those thing and then it is going to feed the results back to the clot. So and after you know as it performs its task it is going to log everything in real time. So it sounds amazing right and uh also we will have one of these files we call it as tools. So you can see this given folders these will be reusable functions like read file write file you know searching the code base or you know running shell or installing dependencies. So by creating these given files this is going to handle all its task even you know running the test also and all the tool inputs will be validated using the zord schemas. So guys this tutorial requires claude models integration. So we could get a response from claude saying uh a JSON response you know for example run this tool with this this input and you can handle that request in the code and return the result and the loop continues again and again. So that we will be executing we are going to understand everything step by step. Now let us try to understand the given project structure. Now if you see this given file sandbox all over here. So this is where all the coding will be done by the agent. Then you can see there is one another file called logs file. So this is going to stores cla tool calls and the result. And you can see I have run uh a project. So you can see there's one file all over here. The next you can see under the src we have one of the another folder tools. So under this we have these respective files fs tool. TS, index.ts, search tool.ts, ts, shell tools.ts, types.ts. So basically here we are going to write all our tool implementation. For example, the shell tools, the FS tools which will be very very much important for this project. Then you can see under the src we have one more folder which is called agent. Now this is the folder which is going to create the brain basically run agent. So that any task you're assigning it, it is going to run and it is going to have the claud's uh connection also all over here. Now you can see I have two JSON files all over here. The first one is package.json and the second one is tsconfig.json. So if you open this file basically what we are doing is we are creating all the important aspects of our application like how do we start you know how we build okay for the development what is the things to be included okay now all the dependencies involved all over here so these will be the given thing now for our given tutorial these are the respective dependencies that we have installed anthropic AI SDK version 0.63 63 then chalk then we have envysa we have fast globe library we have nanoid we have pretty bytes and we have zod and there are some div dependencies like type node ts node tsx and typescript so this will be the given important thing that will be required in our package dojson now you can see we have one another file that is tsconfig so this is going to set up our es modules and type safety things which will be required in the project and you can also see we have a env file which will be very very much important. So you can see all over here we are going to store our API key which you are going to get from the clots platform and also the model name the model version which you are using. So for this project we are using claude sonnet 4 20250514. Now this was a brief introduction for this given project structure. I hope so it's clear to you. Now let us set up our projects. So guys as you can see I have already installed my node and all those thing. But in order to get a folder structure like this for example to create log file to create sandbox to create you know agent runtime and tools what you can do is first initialize the package.json. So for this you can type npm init y and next what you can do is you can create subfolder under this main folder. So you can see type mkdirp src agent src tool src runtime and then sandbox and logs. So all your logs sandbox and the tools and the runtime agent file is going to get created. Now this is pretty simple step. Now guys the next step is installing all those dependencies. So type npm install then anthropic AI SDK chalk envysa fast glob nanoid pretty bytes and zord. These are the very important dependencies that we'll be needing in our project. Then let us set up our ts.config. So for setting up ts doconfig you need to type npm install / d then at the rate types node ts node then tsx and then typescript so that the typescript code that we are using is completely understood by this given compiler. So these are some of the pretty basic you know uh setup things that we have done. Now before we move ahead let us try to understand how our project is working. So you can see index.ts is our main file. It is going to load the ENV file where our cloud API key is there. Then it is going to create a sandbox if it is not present. It is going to read user provided task. Let's say if you give any task to the given uh you know assistant, it is going to use that. It is going to call the agent from the loop.ts file. It is going to start coding, calling tools or iterating or any task based on what you have given it. So it's like you're creating your own agent. For example, like if I run in the command something like npm rundev create a markdown to HTML converter, our agent will start working on that. Sounds pretty amazing, right? And also it is going to read the task. It was also going to ensure that the sandbox folder is ready. It is going to call our run agent and then the clot can start creating files inside the sandbox. it is going to write the code, run the code and also run the test and it is going to iterate until it's done. So that's our given uh project that we are doing all over here. So in a nutshell now let us start with this file. This is very much important file that is tool.ts. Let us try to understand this. Now the first file that we have created is fstool.ts. So guys you can see we have done the particular import this from promises this from path this from fast glob then sandbox from runtime/path then schemas from type okay so these are the respective thing that we'll be creating okay for now we have written something like this since I have already written it so it's not throwing error but when you start doing the development it might throw you error so these are the given imports that we'll be needing so this code actually defines three file related tools. Okay, the first one is list file, second one is read file and the third one is write file so that clot can use to interact with the file system inside a control sandbox directory. And all these three function are asynchronous. They rely on the fast claw for patternbased file matching and fs/promises for non-blocking file operation. And you can see we are also using zord all over here. So for basically the zod schemas so that we can validate and parse the input arguments. So you can see all over here. Now let us try to decode this list file function. So this function takes a pattern let's say star star or /star.ts or star.md. So and list the matching files in the sandbox and it is going to use the fast glob. Okay. So it is using this basically to perform the matching and ensuring that it works with the nested files and the hidden files also for example the env file that we have created and results a list of matching file paths. Then we have this given file which is the read file and this function reads a specific file from the sandbox. It expects the relative path, constructs absolute path by joining it with the sandbox base and then it reads the file content in UTF format and it returns the files content, its path and the number of bytes it contain. So that's pretty amazing. And finally we have write file function. So this function guys writes the content to a file in a sandbox. It ensures that the required folders exist by calling fs.mkdir with the recursive true. Okay, which means that it is going to write the content to the target file and it returns the path and byte size of that written content. Now these functions guys are very much vital and building blocks of the clouds agent tool set. So it is allowing it to read, create and manage files during task execution all while being confined to sandbox for safety and control. Now the next file that we have is search tool.ts. Now it has very much important function that is called search code base piece. So this function is a part of an AI agent toolkit basically from claude that lets it search through the source code files and find lines that contains a specific keywords or phrase. So technically what's this is happening? So basically what we are doing is we are doing a schema parsing and validation. So you can see we have written conquery glob and then schemas search codebase.parse pars argument and uses a zod which is a typescript first schema validation library which is going to check whether the argument contains valid properties like query what is a search keyword for example o or let's say microservices and then it has something other parameter which is glob so which is a file matching pattern for example let's say you are trying to find a typescriptbased file okay so all of these things are being mentioned Now what is the use of this is that it helps avoid runtime errors by validating inputs before using them. Next is what we are doing is is a pattern based file search basically glob pattern matching. Now you can see all over here what I have written all over here is constant uh files await fglo cwd sandbox dot true. Now this is what it is doing is this is performing the patternbased file search. So FG is referring to the fast glob which is a utility that is matching the file path using the glob patterns. For example, the pattern that I discussed you for example the TS pattern for finding the TypeScript based files and uh it is going to do it. Next uh what we are trying to do is we are creating a variable called hits and this is an array in which I have the path, the number and the text. Now you can see all over here we have declared this variable called hits. Okay. So all the result you know will be stored in this given array. So that when we are doing a for loop to search for these respective files you know basically for example in this context we are trying to read the file content. So you can see here we have added con content await fs do read file full in the UTF format. Now what it is doing exactly? So it is asynchronously reading the file content as a UTF8 encoded string and fs/promises is a modern promise-based file system module in NodeJS. So this makes the entire application kind of non-blocking making the program fast and also efficient. Then what we are trying to do is we are trying to do splitting and searching line by line. So you can see all over here we have declared constant lines content.split. Then there's a regular expression. So we are splitting the entire file content into lines and we are using regular expression for the same. So if you don't know about regular expression no worries you can use the same thing and uh what it is saying that it is saying that it is going to handle unique based file systems and also windows based file systems. Next you can see all over here we have a casebased uh repeat next thing what you can see all over here we have a case insensitive matching. So basically what this is doing is this is converting the both of the lines or the query to lowerase to do a case insensitive match which is very much important and finally it is going to hit the query and it is also returning the hits. So this is what this search toolts is doing. Now you can see all over here of lines for each line do I if line dot to lowerase include query to lowerase hits dot pushpath relative line i +1 text line dot slice 0 to240. Now what this is doing is so for every match it is going to create a object with its relative file path and also what it is going to do is it is going to start from one. Okay. And then what is going to happen? It is going to have a 240 character preview of the matching line. And finally, you can see it is returning query, comma heads. So basically, it is sending a JSON-like object that contains the original search keywords and all the found results. So this is the two uh repeat. So in context of this given file, that's why this given file is very much important. So I hope so you would have got the idea why we are writing this given file. Now let us move ahead and try to understand the shell tools.ts. Now this is a shell module and this module defines three key utilities. For example, run shell and installing dependencies and also running test which allows the AI agent to interact with the system shell environment in a controlled and a safe way to avoid security risk. The function is safe ensures only a white list of approved commands like node npm CAD get to be allowed to be executed. So you can see these are the safe prefixes that we have declared. Now if you go all over here export const run shell. So this is our main function. So this takes a command cmd and uh what is going to do is it is going to you know take it as a input and it first checks if the command is safe or not. So we have created one function function is save now then it executes inside the sandbox directory using excisa. Now what is excisa? Uh it is a modern alternative to child process basically that provides a better promise based control and output capturing and it returns a standard output like um you know with exit code and all those thing so that AI agent can reason about. Now the next function is install depths function which is specialized for installing dependencies. It parses the list of packages, constructs an npm install command using npmi and runs it inside the sandbox. Now this enables the AI agent to dynamically add you know required NodeJS libraries while generating the code. And lastly it runs test triggers like npm tests silent a common script defined in the package.json. Okay, I'll show you to quietly run the automated test basically using chest library. Now, this helps the agent to verify whether the generated code works as expected or not. So guys, all these three functions that we have discussed relies on XISA for clean shell execution and also operate strictly on the sandbox directory to isolate the activity and maintain file safety. So I hope so guys would have got a brief idea regarding this file. Now let us look at types.ts. Now this file is responsible for defining the data validation schemas for each of the tools or commands that the cloud coding agent can run on. So basically it uses the zord library which is a powerful typescript first validation tool to strictly define what kind of input is expected for each tool. For example, the read file schema expects a single string file called path that points to a file inside the sandbox. Similarly, there's a list file schema that allows an optional pattern which is defaulting to this route and it uses a glob syntax to match this files. Now, for the commands like runshell, the schema ensures that only a single line of string command and a current working directory are passed and installing dependency is going to check that at least one package name is given in an array. And then the search code base takes a search string basically a query and a glob pattern defaulting to files like search.js or ts files and finally it is going to run test that doesn't expect any input at all. So guys its schema is just an empty object. Now this file ensures that before any function like run shell read file etc is executed the arguments that are passed are validated. If invalid data is passed in, it's caught early and won't crash the app so that you know any future bugs don't come up. Now this is very much important in AI agentic workflows where AI might generate unexpected or incorrect inputs and zod is acting as a safety net all over here. So I hope so guys you would have got a brief idea regarding the functions used inside the subfolder tools. Now let us explore the runtime subfolder. So in which we have a file called path.ts. Now this is this file is defining and all the exports and important constants that are used throughout the claude agentic AI project basically to manage the file path. So first it is importing path. So this imports NodeJS built-in path module which is used safely you know to join or manipulate file path across different operating systems like Windows, Linux or Mac OS. Then you can see in the next line we are exporting the root process. CWD. So this is returning a current working directory from where the NodeJS process is running. So this becomes the base path for your entire project and all other paths are relative to this. Then you can see we are exporting the sandbox. So what this is going to do is this is setting up a path to a folder named sandbox inside our root project which is claude agentic. And this is where the AI agent is going to write the code, create files or execute commands in a safe and isolated environment. Think of it like a playground for the agent. And all the read and write operations are restricted to this folder to prevent any unintended changes to your real project files. And finally, you can see there is a log. So we are exporting log. So this points to the logs folder in our project directory. It can be used to store execution logs, debugging information or error traces. So this is where its importance is there. Now let us move ahead to the agent file. So in this we have two files named clot.ts and loop.ts. Now if you open the cloud ds file. So this file defines the core claw chat handler function which is used in our coding agent project. So you can see we have uh created something like this conclic and here we are extracting our API key from the env file. So this line is initializing our you know API key which is fetched from the given file which is in the ENV and then we are trying to use the model which model we are using. So basically we have given this model or it is using uh if it's not present then it it can use the older model also. Then uh we can use the async function all over here. So what this given function it is doing is so it is defining the certain things. So guys this is exported async function in our core interface for sending prompts and receiving replies from cloud. So it takes a list of structured messages as a parameter input and typically in combination with user and assistant message it is able to pass them to cloud. Now for example you can see here we have given like what are the maximum tokens what is the temperature and then here it's written the directives like you are in agentic coding assistant work inside the sandbox directory only follow the tight loop plan called tools and run the test. So this is the given u you know the directive or the workflow that we have created for this agentic AI and then finally we are reflecting and moving towards the next step. Then it uses a JSON tools to cause exactly what's needed. So one tool per possible and then we have also given a very important directive that uh not to fabricate the file contents means not to change anything if it's if it's there and after each step what you have done explain what changed and why. So that seems like a very decent workflow for our this project. Then we can see we have this one another file which is called as loop.ts. Now in this given file what we have done is we have created a very important function which is called as run agent function. Now this function is going to let claude take control of the coding task by receiving your task. For example, you say create a todo cli app and you know from there the agent is going to take the thing and it is going to ask claude to think about what to do next. So it is basically letting claude call tools like read, write files, run shell commands, install packages etc. and it is logging every step and finally it is repeating until the task is marked as done. So you can think this file like a smart loop where claude plays a role of a selfdirected developer. So you can see in the step one it creates a log file await fs mkdir logs recursive true then it logs the path okay and then it creates logs/folder if it's not already there basically it's preparing a file to save every conversation step with clot then it is starting the conversation so in the step two you know what we are trying to do is we are going to create a variable called messages and we say it as internal message and Let's say the role be user and we discuss about the content. What is the content? What is the task? So it works under only the sandbox directory that we have created. So let's say claude is given any simple task like for example the instruction and all this file operation must stay inside this. Now in the step three it is looping through each thought or action. So here cloud gets the maximum step tries to complete the given task. So that you can see all over here we have applied the loop and in the step four what we are trying to do is we are trying to filter the messages for claude. So you can see I've created a variable const filtered message. Mn where claude message equals to m.troll user and m.troll assistant. So this means that claude only accept messages from user and assistant roles only. And these internal messages are like tools that are kept out of the chat button file. So you can see we have imports all over here. Import tool registry from tools. Then we have import chat with claude from claude file. And then we have fs from fs/promises. Then we have path from path. Then we have logs from runtime path. And finally we are importing chalk. Now let us recap once again what tool registry is for. So tool registry is a collection of available tools your agent can call and you can see chat with claude is a function to talk to claude ln and if you look at fs fs/p promises this is a file system function like for example mkdir append file etc but these are promise based next one we have path so path is a utility to work with file paths then we have logs and you can see it is a directory path constant where logs will be saved. Then we have chalk. So these are the colors console output like for example your console can be like cayenne yellow green etc. Now let us explore all over here. So we have type cloud message role user assistant and content Annie. So what does this mean? So this means that only roles that cloud can accept is the role user or assistant. So it is defining the role that cloud can accept. So each message also has a content. So you can see all over here this is extra ro tool for internal loop and this is not sent to the cloud. So every message has a content also. Now if you see this code block this represents a tool request that claude might output. For example type tool use name read file input let's say it could be ID and let's say 1 2 3. So it could be in a format of JSON and it is like uh basically send to the claude. So let's say for example and this is send to the claude. Now let us come to the main part of our function that is run agent function. Okay. Now this is our main function. So you can see we have exported a sync function run task string maximum steps 12. So there will be 12 iteration. Now we are doing a promise all over here. So what is this all about? So basically it is creating the log folder if it does not exist and it prepares a log file with a given time stamp name and u you can also see there is initial message all over here and in which we have the role user content your task work only under the sandbox. So it is giving a directive like when it starts a conversation with claude the first message tells claude that a task and it restricts it to the sandbox folder only. Now you can see we have a loop all over here. So each loop represents one claude response plus possible tool execution and then you can see we are filtering the message for claude. Now you can see all over here at this time it is trying to check for tool request. So constant tool call response dotcontent find C dot any C.ype and basically it is giving a strict equality for the tool use. So cloud response can include like structured block basically for the tool use and this is extracting the first tool request if present and if no tool is requested you can see there is a negation call all over here. So it means that if Claude just replied with the text only no tool call. So it is going to combine the response into plain text and print it in cayenne. So can is a color for the console and it is also going to add a response to the message and if cla said it is done or task completed then the loop is going to stop otherwise the loop is going to continue. Now at this step what it is happening is it is trying to extract the tool name and input from the block and then it is printing yellow which the tool is called with what is the input and finally it is running the given tool. Then you can see we have the exception handling done all over here. So there is a try block and um in this try and catch block what we are trying to find out is that it looks up the function in the tool registry. If found it is going to run it with the input. If not, it is going to throw the error like it is going to throw an error object and finally it is going to show the result. And you can see it is printing the result in gray and finally it is going to save the given state. Now if you see all over here so what is happening at this stage? So at this stage it is adding claude response to the messages. it is going to add a tool message with the result so that claude can see the tools output in the next step and finally the loop is ending. So in nutshell if I have to say after the loop is finishing or the task is done it is going to print a success message and also it is going to show where the logs are saved. So in nutshell this code run an agent loop. It is sending the message to the claude and if the claude replies with the text it shows it and if claude is requesting a tool it is going to run it and log the result and feed back to the cloud and it is going to repeat until the task is completed or max steps are reached which we have included 12 all over here. So this is all about our log file. Now let us try to run this project. Now guys let us try to run this project. So you have to type npm rundev and after that you are going to see output something like this. Your cloud agent would have started working and you can see in the step one this is iteration one. It is going to create a node JS CLI tool which is basically converting CSV to JSON. So this was the target or the task which I gave to this agent and you can see it has started planning and um then it is starting the function calls and it is doing all these activities inside this sandbox folder and um it has also created a read.md file all over here. We are going to explore it and then you can see it is trying to do all these things and u it is making a CLI script to create a NodeJS CLI. Then it is trying to do the testing also. And u so these were the function calls. Finally it is installing all the dependencies by itself. And um let us go to the end and and you can see it is trying to test the given thing. So it is sending the given file all over here and this is a CSV file and you can see there are certain dependencies it has downloaded like CSV parser and just for testing dev dependency and finally all over here this is a step two it is trying to iterate and testing the given thing whether it's working or not and it has done a lot of thing by itself and finally it is giving you a snapshot all over here that test suits one past wrote it as four. And you can see the time it it has taken to deliver this task and um also you can see that this is um you know what it has done is that it has regained the data. You can see all over here based on the given file it has given. So it says it has converted a CSV to JSON format and uh again there is a reiteration happening. So it is going to reiterate until and unless it is you know getting the exact results. So let us go till the end and you can see it has done all those things. So in the step three only it has given us rest of the iterations were not needed. So it has passed all the test. The CLI is working perfectly with this data. Error handling is also done. It is helping system display properly and file output is also working properly. So guys let us now explore the read.md file and uh let us see now if you can see the read.md file here the agent is going to do all the task. Finally the given thing is completed. So guys this was all about our project doing the agentic coding with clot. I hope so you would have enjoyed this video and if you like these kind of videos then do not forget to hit the subscribe button and click the notification bell icon so that you don't miss out any video.