That’s a really interesting handful of problems, Denis. A lot of these are valid frustrations. But the question is… whose fault are they?
The big API bills? Yeah, no joke.
Using Claude Sonnet or OpenAI o1 full-time will cost you anywhere from $2,000 to $6,000 a month.
Mainline DeepSeek? About $300/month.
Groq? You’ll get the same inference volume on some very capable coding models for about $200.
There are cheaper and even free options—running open-source models locally, using tools that offer free inference time. But not all models are created equal. You have to vet them. And it feels like… Denis hasn’t done that.
So where do you start?
1. Find a reasoning model with a big context window (120k+ tokens) and multimodal capabilities. You want cross-domain reasoning and well-developed thought chains. Models with visual reasoning are often (but not always) better at writing code.
Try this test:
“Hunter S. Thompson and Ada Lovelace are collaborating on FizzBuzz in TypeScript. What does the code look like?”
If the result’s boring? Don’t use that model for coding.
2. Set a real system prompt.
Be explicit about what good code means to you.
You wouldn’t expect a human dev to be a mind reader—why expect a machine to know you want OOP with MVC, observers, event buses, and singletons? Spell it out.
Also: remind the agent QA exists. Tell it that QA's standards are brutal. The simple knowledge that the code will be reviewed will change both the code and the general behavior of your agent.
3. You need an agentic framework that doesn’t suck.
Good luck. Most of them do.
A few might meet your expectations, depending on your use case and pain tolerance.
4. Stop trying to build monoliths!!!
Monoliths are dead.
I keep trying to remind people that we're well into the microservices era, but people still want to design their apps poorly. An AI will absolutely let you do that, without a fight. But, AI will absolutely nuke your app at scale if you give it enough rope and vague instructions.
The Better approach?
Build small, self-contained, API-focused microservices, and disposable UI's.
Modular.
Straightforward. Stable.
Hope that helps.