Talk to code, not to people
Getting output your program can trust.
In an agent, the model's output goes to code, not a person. Code needs an exact format.
Extract the student's details as JSON with name, college and year: 'hey this is Rohan from VIT Pune, 3rd year CSE, want to start DSA'
What the model returned
```json
{
"name": "Rohan",
"college": "VIT Pune",
"year": "3rd year"
}
```Your code: JSON.parse(reply)
Crashed
SyntaxError: Unexpected token '`', "```json { "... is not valid JSON
Real data from gemini-3.5-flash-lite, 2026-10-16
Asked politely, the model wrapped its JSON in a Markdown fence and json.loads failed. Given a schema, it returned clean JSON, and “3rd year” became the number 3.
Your code parses the model's answer. What's the most reliable way to get valid JSON?