When an AI Coding Agent Gets Expensive: What I Changed

I built my crypto trading bot with Replit's AI agent, without a developer background. The agent wrote the code, set up the dashboard, and got the bot talking to the Bitvavo API. That part worked well. What I did not fully grasp until I was deep into it was how the billing for that kind of work actually adds up over time.

How agent billing actually works

An AI coding agent does not charge a flat monthly fee for unlimited changes. It charges for the effort it puts in: the time it spends reading your code, thinking through a fix, writing new code, and checking that the change did not break anything else. A small tweak, like renaming a variable, costs little. A bigger request, like adding a new alert type and changing how the dashboard displays it, costs more, because the agent has to understand more of the project before it touches anything.

This is effort-based billing, and it makes sense once you see it from the other side. The agent is not just typing, it is reasoning about your entire codebase every time you ask for something. The more context it needs to load and the more steps it has to take, the more it costs. For a bot that keeps growing with new features and small fixes, those costs are not one-off, they are ongoing.

What changed when I moved to Claude

I kept the bot running on Replit, but I moved the actual code changes to Claude. The bot still lives on Replit's infrastructure, runs there around the clock, and keeps sending me Telegram alerts. What changed is where I ask for new code or fixes. Claude writes the change, I paste it in, test it, and deploy it myself. I lost the convenience of the agent doing everything end to end inside the project. I gained more control over when and how much I spend on a given change.

The trade-off is real. An agent working directly inside your project can run tests, catch its own mistakes, and iterate without me copying code back and forth. Using a separate assistant for the coding and a different platform for hosting means I now carry more of that back-and-forth myself. For someone comfortable reading code, that is a fair exchange. For someone who wants a single tool to handle everything, it adds friction.

When an agent is still the right tool

I would still use an agent like Replit's for getting something off the ground. Going from nothing to a working bot with a dashboard, without writing code myself, is exactly the kind of job an agent is built for: broad, one-time, and worth paying for because the alternative is not building it at all. Where it gets expensive is in the maintenance phase, when you are making frequent, smaller changes to something that already exists. That is a different kind of task, and effort-based billing makes you feel the difference between the two.

If you are weighing whether to start a project with an AI coding agent, it is worth reading how the billing and workflow hold up in practice before you commit. I wrote up what I found in a review of Replit, and if you want to try it yourself, you can sign up here.

I run into the same question with the other projects on my desk: how much should run on autopilot, and how much should I still do by hand to keep costs predictable. If you want to see how I think about tools like this more broadly, I keep a running list of what I have tested at .