Implement the agentic loop (LLM call -> tool execution -> repeat) with: - Confusion counter: tracks consecutive tool failures, signals Blocked at threshold - Token budget tracking: cumulative token usage across turns - Token budget warning: injects 'wrap up' message at warning threshold (default 80%) - Token budget exhausted: returns TokenBudgetExhausted outcome when budget exceeded - LLM failure threshold: signals Blocked after N consecutive LLM errors - Turn limit: stops after max_turns with partial completion - Configurable error handling: tool errors returned to LLM for self-correction Verification: - All 6 agent runtime tests pass - All existing tests still pass (118 total tests passing) - No lint warnings Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>