Transcription
Hello! In this video, I will review the brand-new and much-hyped Replit Agent. For those who don't know, Replit Agent is an autonomous system for development using agents, and it is maximally autonomous. No other agent, neither Lovable, nor Bolt, nor any other agents that develop applications, has this level of autonomy. Replit Agent is something very new. You can throw Replit Agent into your TZ (technical specification), set its mode to maximum autonomy, and just go to sleep. Wake up in the morning, and the application will be ready. It will plan it itself, build it, refine it, and test it using the built-in Playwright browser. It is claimed that it can do it completely autonomously. Yes, it won't get API keys, some access, secrets, or passwords for you. All that is understood from you. But within this review, I wanted to do something useful, not so much for myself personally, but for a client who constantly needs to update their knowledge base, and their managers store all this in Google Sheets, in our favorite ones. So, I want to create an application using Replit Agent where I would get access to all my Google Sheets or any Google Sheets through a Google Service Account. And I could edit these sheets here, meaning change something, save the text here. And when I have edited this knowledge base, I can send it all to the Qdrant vector database in this way. Create a new collection, choose what will go into the metadata, click "Start Embedding," and it will be embedded without any N8N. It's just such a tool, where you can polish it up a bit, make it presentable, and give it to a client who constantly needs to update their knowledge base, and even for money. Enough talk, let's get straight to business and go back to the very start and begin developing this application from scratch. I have compiled a technical specification, I used Google Gemini for this, dictated a prompt to it, and this is the PRD, Project Requirement Document, that I created. And we move on to Replit. Here, I probably won't choose any specific mode, I'll just paste my PRD. I think the agent will understand what I need itself, whether it's a web app or a data app. Well, it's obviously a data app for me. And let's launch it. Honestly, I've only used Replit fully once, I built an application, but that was quite a while ago, almost immediately after they launched agents. And I didn't really like it, there were a lot of bugs, credits ran out very quickly, and I didn't manage to understand anything. And, well, I kind of abandoned it. Now I hope that with all this hype, Replit Agent 3.0 is super, super, super. Let's see. I still think it's a good thing, but it's one thing to hear and read, and another to do something useful yourself. And this application is useful to me because I have one client to whom I can implement this application, this functionality. Replit offers me to start with the application or with the design. 20+ minutes for the application and 5-10 minutes for the design. Well, I really want to click on design, but okay, let's, like seasoned professionals, start with the backend, I guess, and choose the application. The autonomy level can be adjusted. Here, high, medium, low, max. Let it be high, it will probably show some signs of life sometimes. But there's not much to see here, folks. Probably for 20 minutes it will just work silently. Maybe it will ask me something, which is unlikely. I think as soon as there's some result here, I'll immediately use the magic of editing and show what we get. Replit Agent has reported that it has already done a certain amount of work. Now I'm being prompted to enter my secrets. Also, a Google Service Account Key is required, and an OpenAI API Key is required. Now I'll get all these keys, paste them in, and click "Continue" here. We'll keep watching. I entered my API key, clicked "Place." Someone is restarting, I haven't seen anything yet, I don't know anything yet. There's a "Preview" window here and there's also a database. I don't know about the database yet, but there are already tables here, 5 tables. Users, Settings, Project, Project Data, and Embedding Jobs. Well, let's see how it works. Now I need to enter my Google Service Account Key. I'll do that now. It's not that fast. I've entered my key from my Google service account, where my necessary tables are located. And the work continues. I think I still need to add a connection to the Qdrant vector database for credentials. It hasn't asked yet, but I saw that it installed packages, installed the Qdrant Client. I think it will struggle a bit longer. All we can do is wait for this boring pair programming. For the first time, literally a few seconds ago, the application launched. I see that it didn't connect to the Google account. Something is wrong with my key. Therefore, no tables were found. But without tables, we can probably use something. Here's the connection to Qdrant. And it wrote it incorrectly. OpenAI. Strange, why did it ask me for keys? Connection failed, connection failed, and it says "Google ready to use," well, strange. Dashboard, I wanted to see my Google Sheets here. It's launching, I understand, there's some kind of Playwright built-in here, and it's probably going to test this application itself now, and we'll see. I didn't copy the whole key, I only copied this part, the "ending key." And maybe that's why it didn't recognize it, but I think we'll talk about it now, because in fact, I only sent one single prompt at the beginning and did nothing else. Let's expand to full screen. Interesting. Let's see. Here. It clicks with the cursor. Settings. And what? It's quite slow, but it's thinking, apparently. Here, it clicks. It fills in fields there. Enters the API key. Interesting, where does it get it from? Test Connection. So, it's very cool, guys. It built me an application from one prompt. It opened it itself, it's testing it itself. And it's quite interesting. I think I will definitely use Replit Agent. For some of these small, simple, applied things, for which hands might not reach, but still, some small tools are needed. And I think it's very convenient. Give it a task, say, "Go ahead, build it, and when it's ready, let me know." And they also have their own app for the phone. I don't know if you can do full pair programming there, but most likely you can, but at least. You can check the status there, maybe write something in the chat, and it's very convenient. I think this is boring for now, how slowly it clicks, we'll skip this. And we'll reach some next milestone, where at least the tables will appear in my application. I just think that the whole problem is with the API key. These are the kinds of errors that happen sometimes. Well, about 20 minutes have passed. Since I started, it's still not stopping, I still haven't written anything. It's trying to figure out the Google Sheets API now, saying what's wrong with Google, throwing errors and figuring it out. Well, we'll just wait and see. And I also saw that the Google credentials are not correct, but maybe it hasn't gotten to that yet. It reported to me that everything is ready, more or less nothing works, only the OpenAI test works, it's connection failed, quadrant connection failed, and Google Sheets, in fact, there is no integration at all. Well, let's figure out what's going on with the API keys. Let's see. So, I click "Continue building," "working," I haven't written anything yet. Well, I suspect that my service email is entered incorrectly here, because this is a complete mess. These could be some fake data. I don't know where they are from at all. What is being offered? Database, authentication, and so on. It's not starting from the right place at all. That is, the core functionality doesn't work yet, but it already has "Batch Row Operations," what not. Well, autonomous mode is enabled, let it work a bit longer. I hope my credits will be enough. I have a Core subscription, they give some minimal package, at least for this application it should pay. Well, let's see. In any case, I will bring it to the end, so watch this video until the end. It will probably be especially useful for those who use Google Sheets to upload their assistants' or agents' knowledge base. And sometimes you just want to give the client some tool, not Google Sheets, so they don't break anything. You want to give it to them, but at the same time you don't want to waste time, because not all clients are willing to pay for it. But since I have such a client who is willing to pay, that's why I'm bothering and decided to build this thing in this pair programming mode. I am 100% sure that Replit Agent uses Claude, or Sonnet, or Opus. I know it like the back of my hand, I've used it for so many months. And I figured this out from this phrase: "Thank you for the detailed review. You are absolutely right about these critical issues. Let me fix them systematically." This is 100% a Claude formulation. As soon as I read it, all my doubts disappeared. Perhaps a mix of models is used here. Maybe Claude Sonnet is on standby. Maybe some other model is orchestrating it. But it's interesting. It still doesn't pull up the tables. It still doesn't work, but it has somehow figured out that the problem is with my credentials. It writes here that it needs to fix all this somewhere. It's strange that it needs to be fixed, because, at least, it should ask me for them completely to put them into secrets. Well, let it fix it, and we'll see further. Here's Claude, his handwriting, a characteristic feature, "ready for production." Nothing works, nothing connects, and all this is "ready for production." This is not going to work, of course. Let me paste my keys now and see. My keys are already in. Test connection. Qdrant, we didn't connect. Google, we didn't connect. I don't know. Let me paste them again. Save Configuration. Test. Oh, Connect. Well, at least OpenAI connected, that's good. Let's put the Qdrant key in, test connection, and it doesn't work. It wrote to fix at least the connection to Google Sheets and Qdrant connections. "You're absolutely right." Ah, I canceled my subscription to Cloud Code because Cloud Code has become simply unbearable. It writes terrible code, it... All of this has turned into fixes of fixes and fixes of fixes of fixes. In general, it's just complete nonsense. I don't know what Anthropic did with such a magnificent model as Opus, but what they did is unacceptable. And I'm switching to OpenAI Codex for now, I'll work with it, and we'll see what Anthropic does next, but it's obvious that Grok 4 Fast works great, and Codex also works great, Qwen. In general, there's plenty to choose from, and the thing is that Cloud Code used to work phenomenally. Now it's some kind of pumpkin. I was able to connect to Qdrant and OpenAI. It writes here and here. And that's good, but Google Sheets still doesn't work. I updated the credentials. It turned out that it needed not only the key itself, but the entire JSON. But I don't know, it hasn't worked yet. Most likely, I'll need to restart the app for this value to be pulled from the environment variables. For now... And now it's going to restart, it will restart, and hopefully, maybe in the exit run it will work. In essence, this will be the final moment in terms of integration with other services. Then we'll look at the logic. But I think it took a long time to fine-tune the connection to Qdrant. I think it should work in the end. There's nothing complicated in its API, in principle. The application loaded, and all my Google Sheets miraculously appeared. And that's very good news. The bad news is that I have absolutely no idea how to use this application. Here I have some mock "new project," there's no button to create a new project. Nothing. It's unclear. I complained about this, I said, "Listen, my dear, how do I use this miracle application at all?" So it went to think. Ah, so it wrote me instructions. Select a Google Sheet, wait for the sheet to load. And where did it put it? This is just a brilliant user interface. These are pages in the sheet. Okay. I select pages, and here's the "Create Project" button. Okay. Let's create a project. Cool. How? And where is my sheet? Where is my data? Well, synchronize data. Okay. Oh, it appeared, look. Cool. So, I connected Qdrant, I connected OpenAI, so that the model for embedding is available. And I'll try to click the "Embed to Qdrant" button for the first time. I can create a new collection. Replace App Test 1. Reboot. Okay. Start Embedding. And we'll see the progress. I'll open Qdrant now. Ah, everything broke anyway. But nevertheless, Replit Test. Let's see if at least the collection was created. I see that the collection was created. But nothing got into it. And in general, well, I think maybe there's some problem with OpenAI or something else. But it says it fixed it. Do we believe it or not? Of course, I believe it. And again, I'm testing the metadata, start embedding. Oh, it works. Now it works. So, Replit Test 2 collection, right? It's already created in Qdrant. So, what about the metadata? Where are my metadata? Apparently, they are being transferred incorrectly. Well, that's fine, these are all minor things. The main thing is that the basic history works, even if it's super slow, it works. And this means clients can update it calmly. So, you can add tracking functionality to re-embed only those chunks that have actually changed and quickly create some authentication for this application, and it will be cool, in my opinion. Replit Agent has made a very cool thing, this is a true pair programming tool, it just gave maximum autonomy, set it up for the night, and it made something simple like this for you quite quickly, literally with 5-10 prompts and some useful tool like this, even if it's simple, can be made. I am very, very pleased with this. Somewhere it lags, somewhere maybe Claude lets it down a bit. I highly recommend you try it. If you haven't tried Replit, especially after the latest update Agent 3. This is a really cool pair programming tool for those who don't know how to program. If you've watched this video to the end, I thank you. Please write comments, what you think about pair programming, have you used Replit, ask me questions, I'm always happy to answer them. And constructive debate, constructive criticism, I always welcome. Like, subscribe to the channel, hit the bell so you don't miss my next videos. I'll have quite a few more. I try to talk more about automation, about using artificial intelligence in development. See you in the next video. Bye! Subtitle Editor A. Semkin Proofreader A. Egorova