GPT agents failing on ambiguous input — how do you handle it?

[JSONFIRST]

GPT-4 is great but still breaks when users say things like 'do the thing from yesterday' or mix multiple intents in one sentence.

We use JSONFIRST to pre-parse input into { action, object, confidence } — then only call GPT if confidence is above threshold.

How do you handle low-confidence or ambiguous input in your OpenAI agent pipelines?