Build an AI app over ten days
For the past 10 days I've obsessed with Replit Agent. Task: For the past 10 days I've obsessed with Replit Agent. I've been building an AI-based app that I've always wanted but didn't want to spend the time and money to actually build it. Anyhow, I've actually built the darn thing and it works better than I expected. Note, I'm pretty good at prompting (I read a lot about prompting skills and have been heavily using ChatGPT since it's first public release) so my experience may be different from those who expect to get a polished product from just a few prompts and are not willing to put in the work. PROS: Proactive feature additions (also a con): Agent added features to my app that I didn't ask for but made all the sense in the world - that's the amazing part. It has a better UX because of these additions. I asked it to build an admin page for my app and with few instructions it did a damn good job that I was able to improve on. Creativity: Give it a good prompt that explains generally the functionality you want and it takes care of the rest. Gets you 80%+ of the way there. Access to the code: You can manually tweak the code and UI if you're so inclined. Or if you don't care about spending $.25 for each check-in, just tell the Agent what you want changed and it'll do it for you. End-to-end development: It creates the app, connects to the back end and deploys. All databases and fields required to support your app are created on your behalf. Log in capabilities are easy to enable. CTO capabilities: I often ask the agent advise on system design, how I could make the app better, what it's weaknesses are, and suggestions for features/functionality and it usually provides really insightful responses. Examples, it gave me advice on what I should and should not prioritize for my MVP, compared my app to others on the market, and gave me advice on what Unlocks my ideas: I have tons of ideas for personal and marketable apps. The obstacle was always the time, money and skills to create them. I'm feeling unlocked now. ****REPLIT TEAM****: BIG THANK YOU. I APPRECIATE YOU GUYS!!! CONS: Too proactive: Too Agent often does something before you give it official approval to do so. You may be thinking through some ideas with your prompt, but it will take it as an instruction to do it. I've told it many times to ask me for approval before implementing things (I even asked it to save this command in the project .md file), but it "forgets" and will do it again at some point later. Debugging loops: Sometimes when I tell the agent to fix a tricky bug, i see in the progress tracker posts that it's doing the same things over and over. Often it says "I found the issue!" and says it's fixed, only for it not to be fixed or the problem be made worse. Then I start another round of fixes. These "fixes" get expensive at $.25 a pop. ****REPLIT TEAM****: Can you implement a system where we only get charged for one round of fixes for the same problem? If the agent has to try 3+ times for the same problem, that's not our fault. Testing: When working with UI changes, the agent often says "I tested XYZ and it's working. Please try it out" but sometimes I try it out and the page doesn't even scroll or the button click doesn't even work - basic stuff. I spend too much time with these UI basic bugs. ****REPLIT TEAM****: can you enable a feature where the agent can actually interact with the apps to test basic UI functionality? Check-in costs: At $.25 per checkpoint, building an app can get expensive. I admit that the alternatives -- pay a programmer, build with no-code tools like Bubble, seek Q&A help on the internet, etc. -- are often much pricier and time consuming. OAuth Authentication: The Replit OAuth authentication can be buggy / problematic. In summary, Agent has its issues, but overall I'm a fan and am excited to see how the Replit team improves the product in the coming months/years. EDIT: I'd also like to see... Rate limiting / throttling for dummies: A simple way to implement throttling and backend API rate limiting. I know about and was able to implement this stuff with prompts because I'm well verse in tech, but for the average person who's building an app they want to expose publicly, they may not even know they need to protect their apps from fraudsters/hackers who will try to exploit their app.; tools: Replit Agent. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include strong prompting skills and willingness to review code. Operating costs are significant at $0.25 per checkpoint, especially during debugging loops. Integration uncertainty exists regarding OAuth authentication stability and the need for manual implementation of rate limiting for security. Human approval is critical to prevent the agent from executing unapproved changes..
Tools
Published on Agent Work Atlas:
Overview
For the past 10 days I've obsessed with Replit Agent.
Workflow
For the past 10 days I've obsessed with Replit Agent. I've been building an AI-based app that I've always wanted but didn't want to spend the time and money to actually build it. Anyhow, I've actually built the darn thing and it works better than I expected. Note, I'm pretty good at prompting (I read a lot about prompting skills and have been heavily using ChatGPT since it's first public release) so my experience may be different from those who expect to get a polished product from just a few prompts and are not willing to put in the work. PROS: Proactive feature additions (also a con): Agent added features to my app that I didn't ask for but made all the sense in the world - that's the amazing part. It has a better UX because of these additions. I asked it to build an admin page for my app and with few instructions it did a damn good job that I was able to improve on. Creativity: Give it a good prompt that explains generally the functionality you want and it takes care of the rest. Gets you 80%+ of the way there. Access to the code: You can manually tweak the code and UI if you're so inclined. Or if you don't care about spending $.25 for each check-in, just tell the Agent what you want changed and it'll do it for you. End-to-end development: It creates the app, connects to the back end and deploys. All databases and fields required to support your app are created on your behalf. Log in capabilities are easy to enable. CTO capabilities: I often ask the agent advise on system design, how I could make the app better, what it's weaknesses are, and suggestions for features/functionality and it usually provides really insightful responses. Examples, it gave me advice on what I should and should not prioritize for my MVP, compared my app to others on the market, and gave me advice on what Unlocks my ideas: I have tons of ideas for personal and marketable apps. The obstacle was always the time, money and skills to create them. I'm feeling unlocked now. ****REPLIT TEAM****: BIG THANK YOU. I APPRECIATE YOU GUYS!!! CONS: Too proactive: Too Agent often does something before you give it official approval to do so. You may be thinking through some ideas with your prompt, but it will take it as an instruction to do it. I've told it many times to ask me for approval before implementing things (I even asked it to save this command in the project .md file), but it "forgets" and will do it again at some point later. Debugging loops: Sometimes when I tell the agent to fix a tricky bug, i see in the progress tracker posts that it's doing the same things over and over. Often it says "I found the issue!" and says it's fixed, only for it not to be fixed or the problem be made worse. Then I start another round of fixes. These "fixes" get expensive at $.25 a pop. ****REPLIT TEAM****: Can you implement a system where we only get charged for one round of fixes for the same problem? If the agent has to try 3+ times for the same problem, that's not our fault. Testing: When working with UI changes, the agent often says "I tested XYZ and it's working. Please try it out" but sometimes I try it out and the page doesn't even scroll or the button click doesn't even work - basic stuff. I spend too much time with these UI basic bugs. ****REPLIT TEAM****: can you enable a feature where the agent can actually interact with the apps to test basic UI functionality? Check-in costs: At $.25 per checkpoint, building an app can get expensive. I admit that the alternatives -- pay a programmer, build with no-code tools like Bubble, seek Q&A help on the internet, etc. -- are often much pricier and time consuming. OAuth Authentication: The Replit OAuth authentication can be buggy / problematic. In summary, Agent has its issues, but overall I'm a fan and am excited to see how the Replit team improves the product in the coming months/years. EDIT: I'd also like to see... Rate limiting / throttling for dummies: A simple way to implement throttling and backend API rate limiting. I know about and was able to implement this stuff with prompts because I'm well verse in tech, but for the average person who's building an app they want to expose publicly, they may not even know they need to protect their apps from fraudsters/hackers who will try to exploit their app.
Reported result
The agent reportedly gets the user 80%+ of the way to a functional app. However, it exhibits issues such as acting without approval, entering debugging loops that increase costs, and failing to properly test basic UI functionality. The user notes that while the agent is creative and proactive, these traits can lead to unwanted changes and expensive fix cycles.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include strong prompting skills and willingness to review code. Operating costs are significant at $0.25 per checkpoint, especially during debugging loops. Integration uncertainty exists regarding OAuth authentication stability and the need for manual implementation of rate limiting for security. Human approval is critical to prevent the agent from executing unapproved changes.
Try it
I want to build a simple AI-based web app using Replit Agent. My goal is to create a [replaceable: app description, e.g., task manager with AI suggestions] over the next few days. I have [replaceable: budget limit, e.g., $50] for checkpoints. Please follow these steps: 1. Start by creating a basic project structure for [replaceable: tech stack, e.g., React and Node.js]. 2. Implement core features: [replaceable: feature 1], [replaceable: feature 2]. 3. Before making any major changes or adding new features, ask for my explicit approval. 4. If you encounter a bug, attempt one fix. If it fails, stop and report the issue to me instead of looping. 5. After each significant update, provide a summary of changes and ask me to test the UI. Output format: Provide code snippets for new files and a clear list of steps for me to test. Success metric: The app runs locally without critical errors, and I approve the UI flow. If inputs are missing, ask clarifying questions. Do not proceed with expensive operations without confirmation. Start with a small sample feature to verify the workflow.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.