When you pay $200+ a month for your coding agents, you might fall into thinking that it's gonna do the job. If it doesn't, you could assume it's not possible and do it yourself, losing valuable builder time. Let's look at an example, Claude Code's failure and Codex's solution.
In the past, integrating a third party's webhooks meant you had to first assume the right events for the right state changes in your app. Then, you had to reproduce the entire lifecycle, which was an involved process that needed to be done manually. It's not an option to just look at the docs when you have a deeper integration. They _always_ deviate from expected reality. For example, you might need to be notified about _every_ purchase event in your app. The docs show a payment.succeeded event, but they don't mention if you're gonna get it for the beginning of a trial or a free product purchase. And just looking at the example payloads, it might not be clear how to differentiate between initial subscription purchases vs. renewal payments. Etc., etc., there are many more cases and edge cases, so reproduction is a must. You can give the docs to AI, tell it all the good context and requirements, give it existing examples that use good abstractions to do the right implementation, and you will still get slop. Until you have a record of all possible events. And going through that record itself means a big time investment: Setting up different product types, setting up the checkout, purchasing the products in different combinations (just to stick with our payment.succeeded example).
Obviously, the obvious#
Obviously, you would want the AI to do the setup via API or UI. Which obviously I did. And obviously, you would want the AI to do the purchases itself, which should not be a problem in local dev and a sandbox account, but Claude just didn't want to do it, no matter how I asked it.
In cases like this, the (non-?) obvious weapon seems to be Codex. Whenever Claude Code is behaving weird or lazy in the browser, Codex seems to be the ultimate solution. But here comes the kicker: Codex couldn't do it initially, either. I'm not exactly sure if it's because I'm using Codex and Claude Code through the Conductor app, but this is what I got from Codex, for example:
Blocked after three browser-connection retries. The bridge fails during initialization with Importing module "node:process" is not allowed, so the five card-based checkouts cannot be completed through Whop’s required interactive checkout.
Instead of continuing to dig into the harness setup, I got out the ultimate ultimate Codex weapon: the ChatGPT App. It has this "Computer Use" thing and might be the solution to any automation problem. It went through everything, and more, no questions asked. I had all the products, records, purchases and analysis to review in a matter of some good minutes. (maybe a couple of hours)
The point here is that your vanilla Claude Code can do one thing. Conductor has a bit of a harness on top of it, so it might be able to do or NOT do another thing for whatever reason. And then there are tools like AI APIs and ChatGPT that might be able to tackle problems from entirely different angles. Don't get hung up on just one of them.
Rails.Builders Continuous r-AI-ls.Builders Edition#
This is, I think, the 6th cohort since 2025. I'm really excited about where this one goes. We already have very strong Builders with real businesses and real offers joining the show beginning next week, sharing their journey in live sessions on business and building with r-AI-ls: extremely educational, highly gamified, MINSWAN social, and with some new features (private active-Builders-only Slack channel with Professor Otto Labot). All the info is over at rails.builders — please check it out and get yourself waitlisted (or activated immediately if you are an "OG").
[AIN#34] AI Nuggets Week#34#
📺 Expert Knowledge YouTube Nuggets#
Agents Are Your New Power Users: How Klaviyo CEO Andrew Bialecki Is Remaking a $1.4B Business (SaaStr.ai podcast, Andrew Bialecki, Klaviyo) - If you are interested in big orgs (this one is marketing-related too) getting re-organized to become AI-first, this is a killer disruptive one for your ears.
🎰 Just for Fun links#
Nothing to see here as well. I only really laughed in real life this week. I'm sure next week will be more fun in the www.