Talaria: Sync-Based Headless Coding Agent

Sync-based headless coding agent. Fewer round-trips. Smarter models. Better code. Traditional tool-calling agents make one LLM inference per tool call. A typical coding task (read files, modify, test) requires 6+ round-trips. Each round-trip costs tokens and adds latency. Code Mode flips this: the LLM writes a TypeScript script that orchestrates all the tool calls, executes it in a sandbox, and returns the result in a single round-trip. The agent runs in hybrid mode — it uses direct tool calls for simple single-step tasks and Code Mode scripts for complex multi-step operations. The LLM decides which approach to use based on task complexity. Economic insight: Code Mode's 60-80% reduction in LLM calls means frontier models (Opus 4.6 at $5/$25 per MTok) cost less per task than mid-tier models (Sonnet 4.5 at $3/$15 per MTok) with traditional tool-calling. Better model + lower cost.