Agent-driven development in Copilot Applied Science

What this post added

This post details the development of an agent-driven development workflow using GitHub Copilot CLI and Claude Opus. It introduces the concept of 'eval-agents' for automating the analysis of coding agent performance trajectories. Key contributions include strategies for effective prompting (conversational, verbose, planning-first), architectural principles for agent-first repositories (refactoring, documentation, testing), and iteration strategies based on 'blame process, not agents' to enable rapid development. The post highlights the successful creation of 11 new agents, 4 new skills, and the concept of eval-agent workflows in under three days by a team of four scientists, demonstrating the efficiency gains of this approach.