Where good ideas come from (for coding agents) (sunilpai.dev)

A new approach to enhancing the effectiveness of coding agents, particularly large language models (LLMs), has been proposed by analyzing the concept of "idea-space" through Steven Johnson's framework, "Where Good Ideas Come From." The key insight is that LLMs excel at generating ideas when users provide specific constraints and context, essentially guiding the model through "adjacent possible" trajectories or small, incremental changes instead of vague requests. This method focuses on enhancing user-adaptation by clearly defining goals, which leads agents to produce more relevant results. The article illustrates a week-long simulation demonstrating how introducing elements like context packets, performance feedback, and iterative testing can transform interactions with coding agents into structured engineering practices. By employing techniques such as building "liquid networks" of relevant information and leveraging error as a means of steering progress, users can harness the strengths of these agents to foster innovation while minimizing common pitfalls. This shift emphasizes a collaborative workflow where both user input and agent capabilities can converge to produce robust software solutions.