Really folks. That idea of people without experience and knowledge of software engineering vibe coding production code is a little bit not realistic, no matter how much Claude wants to convince investors of the contrary.
Coding agents are fantastic tools, but they are not Jesus: they don't do miracles.
A big part of the cost explosion is using the frontier model for tasks that don't need to be done. I tried using the gpt-4o and much cheaper models, and the cheaper ones were more accurate in my three — paying for the reasoning depth that I don't use. The other half is asking the model to do what deterministic rules should be. Calls that don't are the cheapest. Starting with profiling what calls the larger model really needs.
Very interesting. I am guessing the company didn't offer refunds if the receiving computer wasn't able to process the results, unless the issue was on the service...
> A deterministic failure doesn’t get better when you retry it
Among all the other nonsense here, this one is solid advice. Most software with retries just blindly retries N times - in reality, you need a retry/backoff policy based on the specific type of error encountered (I've been rooting these out of our own software stack lately)
9 comments
[ 3.4 ms ] story [ 35.8 ms ] threadCoding agents are fantastic tools, but they are not Jesus: they don't do miracles.
Among all the other nonsense here, this one is solid advice. Most software with retries just blindly retries N times - in reality, you need a retry/backoff policy based on the specific type of error encountered (I've been rooting these out of our own software stack lately)