LLM keeps mistaking agent name argument when doing...
# supporta
Ashimpto
05/21/2026, 12:05 PMFor a squad all running gpt 5.4 mini. There are times when instead of doing handoff to "lake elsinore" agent (that's its name, written correctly everywhere in prompt and handoff), it's sending the argument as lake_elsinore, which leads to "agent not found" critical error ending the call.
How can this be solved? Is there a way to pass that argument server side or in any way not text so it doesn't lead to such recomposition of text?
Call ID: 019e2ae7-78c5-7ffd-bbeb-717a1ce6d30a
Date: May 15
https://cdn.discordapp.com/attachments/1506991321955700766/1506991322102366298/image.png?ex=6a104689&is=6a0ef509&hm=08215a9c55b1969e1e0e2e5b15b20b83a2ffc12407293f53010f3127a9daf105&
s
Shaunak
05/22/2026, 5:44 AMHi,
Thanks for reporting this along with the call ID - we were able to identify the issue.
Shaunak
05/22/2026, 5:45 AMThis is a known behavior with GPT 5.4 mini specifically. The model sometimes normalizes spaces to underscores in function call arguments, so when it tries to hand off to "lake elsinore," it sends "lake_elsinore" instead, which doesn't match the agent name exactly and causes the handoff to fail.
Immediate workaround: Rename your squad agents to use underscores or hyphens instead of spaces - for example, CHECK_ACCOUNT_AGENT_LAKE_ELSINORE or check-account-agent-lake-elsinore. That way, even if the model applies its normalization, the name still matches correctly.
Shaunak
05/22/2026, 5:45 AMWe've flagged this internally as something that should be handled more gracefully on our end so that minor formatting differences don't break handoffs. In the meantime, removing spaces from agent names should resolve the issue for you right away.
Let us know if you run into anything else.