Code Slop Isn't a Model Problem, It's a Harness Problem (www.aakash.io)

Recent insights from Wyndly highlight that the challenges in using LLM coding agents stem not from the models themselves but from the surrounding processes, termed "harness problems." Observations revealed that as the context window filled with irrelevant information, the output quality deteriorated—a phenomenon known as "context rot." This degradation can be addressed by ensuring agents tackle single, well-defined tasks without accumulating distracting information. Additionally, challenges such as vague specifications and lack of feedback highlight the need for systematic processes that foster clarity and accountability before coding begins. To combat these issues, a new tool called Oro has been developed, serving as an orchestrator for agent swarms. It employs methods like maintaining strict workflows, enforcing automated feedback loops, and preserving institutional knowledge across sessions. By assigning agents to isolated tasks within separate git worktrees, Oro not only enhances concurrency but also minimizes conflicts. These strategies collectively underscore the importance of structured workflows in maximizing the potential of AI coding agents, reinforcing that the model's capabilities can only be realized with an effective harness.