The PyranaHarness implements native LLM tool calling through LiteLLM by creating a Turn 0 plan, running iterative tool-calling loops, and finalizing an AgentOutputEnvelope so that agent orchestration proceeds from planning through persisted output.

PyranaHarness (pyrana_services/orchestration/harness/pyrana_harness.py) which implements native tool calling via LiteLLM

— PYRANA_Internal_Knowledge_Graph_Specification.md

Turn 0: create initial plan via LLM

— PYRANA_Internal_Knowledge_Graph_Specification.md

Native tool calling loop: LLM emits tool_calls → execute tools → observe → repeat

— PYRANA_Internal_Knowledge_Graph_Specification.md

Finalize: build and persist AgentOutputEnvelope

— PYRANA_Internal_Knowledge_Graph_Specification.md