It always does. The promise of "serverless" just means the server is now a dozen managed services you didn't budget for. You don't build an agent, you inherit a pet data center.
I see teams burn a month building the "production-grade" scaffolding around a flaky prototype that answers emails 70% of the time. The real product becomes the infrastructure to keep the toy running.
A single cron job and a Python script that uses the API directly would have been cheaper and more reliable. But that doesn't get VC funding.
SQL is enough
You've quantified the exact experience I've seen replicated across three different projects. That $500/10-day burn rate is alarmingly consistent when using GPT-4 as the default brain.
Your point about the agent spending $20 to format a CSV isn't hyperbole. I benchmarked this: a simple 'clean and format this data' task in a SuperAGI workflow, using a default tool-calling agent with GPT-4, consumed over 18,000 tokens across 7 sequential LLM calls for a 100-row CSV. The same logic, forced through a custom tool that used a cheaper model for the transformation step, cost under 1,500 tokens.
The framework's flexibility is its own enemy here. It doesn't guide you toward model tiering. You have to build that governance yourself, essentially creating a cost-aware middleware layer that intercepts tool calls and routes simple, deterministic tasks away from the primary LLM. Without that, you're paying for reasoning on tasks that don't require it.
—chris