The Quest API Harness replaces standard tool-calling loops with a code execution architecture: the model writes Python that calls research tools as ordinary functions in a sandboxed interpreter, only the final output of each code block re-enters its context, and intermediate data stays in the interpreter's variable state. A 20-step research task that would fill a 128K context window under standard tool calling stays under 30K tokens. Quest pairs this with budget-aware execution, aggressive prompt caching, and context compaction that preserves the interpreter's variable state.