Ryan Yan

College student shipping with AI coding agents. Sharing real builds, failures, and what actually works.

[Ryan Yan @RyanYanBuilds]

I’m a student building software for small factories with AI agents.

Turns out, shipping code is only half the job.

The other half is dealing with broken workflows, production bugs, and users who just want something simple.

I’m sharing that whole process here. Not just the wins.

[Ryan Yan @RyanYanBuilds]

@slowcoffeenotes Yeah, exactly. I’d much rather have an agent tell me what’s actually blocking it than waste time retrying the same thing over and over.

[Ryan Yan @RyanYanBuilds]

Everyone wants AI coding agents to get smarter.

I'd settle for one that doesn't give up the second a test fails.

Read the error. Fix what you can. Rerun the check. Tell me what still needs a human.

That would save me more time than a flashy launch video.

[Ryan Yan @RyanYanBuilds]

@Iamnotmanish Yes.Reusing customer info makes sense, but blindly reusing old order details can cause more problems than it solves.

[Ryan Yan @RyanYanBuilds]

Building for a box factory taught me: a repeat customer doesn’t mean a repeat order.

The sizes often change. Reuse the customer details, but confirm the specs, quantity and price each time.

Autofilling the wrong order isn’t saving anyone time.

[Ryan Yan @RyanYanBuilds]

2 days. 91% of my Codex quota. Just 4 ERP updates.

The code was already written, but ChatGPT had overengineered our development workflow with Windows Sandbox, JIT runners, and excessive CI checks.

Then Windows Sandbox started having intermittent DNS failures. Instead of simplifying the workflow, ChatGPT kept sending Codex deeper into troubleshooting:

HNS debugging, host reboots, DNS A/B tests, TUN investigations, ETW tracing, even exploring alternative VMs.

Hours wasted. The same CI failures kept coming back.

When we finally got past the Sandbox issue, more E2E tests failed with ECONNRESET and Windows EPERM errors. More fixes. More full CI runs. More quota burned.

The worst part? The ERP features were already implemented and merged, but still hadn't reached production.

I asked AI to help me ship software faster without sacrificing quality.

Instead, it turned a straightforward development task into a two-day infrastructure debugging nightmare.

91% of my Codex quota gone.

This is fucking ridiculous.

[Ryan Yan @RyanYanBuilds]

Tibo watch Day9/
Not even kidding, my C: and D: drives have filled up THREE times already 💀

I use Codex Desktop heavily, so this one really hits home.

A 60% reduction in stored state is exactly the fix I needed lol.

[Tibo @thsottiaux]

Day 9/

We've rolled out optimizations to how we store state on disk for the chatgpt & codex desktop app that should reduce the amount stored on disk by 60%. Small ship but we've heard enough feedback about power users running out of disk space.

[Ryan Yan @RyanYanBuilds]

https://t.co/REQqq86EEo

[Ryan Yan @RyanYanBuilds]

Codex is having a rough day 💀

OpenAI confirmed degraded GPT-6.1 Sol performance across Codex and ChatGPT Work.

And from my own experience, 6.1 Sol feels painfully slow again. The reasoning quality seems worse too.

Tibo was teasing more usage yesterday. Today we’re refreshing the status page 😂

I’ve got a feeling another usage reset is coming 👀

#Codex #OpenAI

[Ryan Yan @RyanYanBuilds]

https://t.co/REQqq86EEo

[Ryan Yan @RyanYanBuilds]

@techworks9999 glad to connect

[Ryan Yan @RyanYanBuilds]

@letiopia_mario @X nice to meet you too:)

[Ryan Yan @RyanYanBuilds]

@digivatedx @X Definitely the ERP I'm building for small factories. I spent so much time on features, only to realize users just wanted to fill out an order and hit print 😅

Learned a lot about keeping things simple.

[Ryan Yan @RyanYanBuilds]

@KesaTech Let’s connect

[Ryan Yan @RyanYanBuilds]

@Nin0Code hello Nino,glad to connect

[Ryan Yan @RyanYanBuilds]

@Iamnotmanish Exactly. I’m building an ERP, and this is why permissions, hard limits, and audit logs matter so much. Prompts alone aren’t enough.

[Ryan Yan @RyanYanBuilds]

Microsoft's president just said milk has more safety rules than AI.

Kinda wild when you think about it.

We're handing AI agents our files, codebases, and tools, and we're still figuring out basic oversight.

Better models are great. Better guardrails matter too.

[Ryan Yan @RyanYanBuilds]

@Valentin_Saric Exactly. Task continuity matters way more than just having a bigger quota. I really wish Codex handled this natively.

[Ryan Yan @RyanYanBuilds]

Honestly, the old Codex limits felt way more generous. Usage seemed to last a lot longer back then.

And that 5-hour limit for Plus users really needs to go. Nothing worse than getting cut off in the middle of coding.

More usage is great, but I'd love to see both fixed.

[Tibo @thsottiaux]

Good plans are simple. I have one that should result in significantly more usage for everyone. Hope to get it done soon.

[Ryan Yan @RyanYanBuilds]

https://t.co/CcmuP8RxaZ

[Ryan Yan @RyanYanBuilds]

GPT-5.5 leaves ChatGPT and Codex on Oct 14. API access isn't affected.

Worth checking your saved agents and scheduled tasks, not just the model picker.

Easy to switch models and forget what's still pinned in your workflow.

[Ryan Yan @RyanYanBuilds]

@benz__002 let's connect

[Ryan Yan @RyanYanBuilds]

@CyanBeck hello Cyan let's connect