Realtime API tool calls causing 500 server_error response
OpenAI Developer Community
February 10, 2026
After more testing it looks like the problem isn’t the length of the users response.
It’s actually : the need for a tool call to be performed.
I’ve also noticed this kind of behaviour has started appearing more often, and is not tied to the first interaction between the Realtime agent and the user.
To prove this I did the following:
* replace current orchestrator & subagent setup with a simple 1 agent setup
* remove all tool calls
The bug never appears in this simple setup no matter how many times I try to reproduce it.
Hoping these additional details will help somebody aid me in sorting this out.
Discussion in the ATmosphere