Learning path · Tools & MCP · 69
Function Calling
Vendor API pattern where models return named functions with arguments matching predefined schemas for runtime execution.
Why it matters
- Standard integration path on OpenAI, Anthropic, and others.
- Enables parallel tool calls on supported models.
- Non-deterministic JSON serialization affects caching.
Key ideas
- JSON schema parameters
- Parallel calls
- Tool choice modes
Function calling binds model outputs to typed functions your runtime dispatches. Use explicit enums and required fields; optional sprawl confuses smaller models. Handle partial failures when parallel calls return mixed success. Serialize tool definitions in stable order for prompt-cache hits. Log arguments and results for replay debugging. Publish tool latency SLOs separately from model latency; users blame "the AI" when CRM lookups stall for thirty seconds. Ship only after eval gates pass on representative production failures.
Updated 2026-08-09 · Full learning path