Design JSON Schemas for LLM function calling that hold up under real production traffic. Strict mode, enums, nesting depth, Pydantic-first, and the failure modes I measure across OpenAI, Anthropic, and Gemini.
Parallel tool calls cut agent latency by 40-70%, but the OpenAI, Anthropic, and Gemini APIs behave differently. Runnable Python for each provider, the opt-out flags that matter, and paired-eval methodology for deciding when to serialize.
When to pick structured outputs vs function calling across OpenAI, Anthropic, and Gemini in 2026: schema enforcement, latency, and runnable code for each pattern.