Run an AI-first development loop (scope, plan, delegate to coding agents, verify, ship), preferring agent execution over hand-writing whenever quality holds, and build and maintain the team's agent harness for development itself (project memory, skills, hooks, MCP, CI checks) so agents can safely do bounded engineering work. Evaluate the agent end-to-end on real multi-step tasks, then debug its behavior by reading traces and finding root causes, fixing the harness, tools, or prompts to improve reliability, tool-use accuracy, cost, and latency.