This may be due to the way you’ve included your workflow example here, but your agent doesn’t have a model attached. Without a model, it can’t reason about tool usage.
I’d recommend trying to improve the prompts a bit as well - both the system prompt as well as the tool description. These use slightly ambiguous language that could be difficult for the model to understand. I’d recommend explaining your use case to a LLM and asking it to write tool descriptions and a system prompt to the agent.
Is it possible to debug what kind of schema was actualy generated, just to have a clue where the issue?
Maybe it’s better to try generate full json as output and then pass it to the http request node (as distinct node, not tool)?
PS: it’s worth to mention - I have the error not all the time, 50% of workflows are succeeded. Sometimes even the same input may lead to success, but sometimes leads to error