Reddit Developers Discuss How to Optimize Harness Architecture for Long-Task AI Agents

On November 7, 2026, a thread appeared on Reddit's LocalLLaMA community in which a developer shared and asked for advice on improving the execution efficiency of a custom AI agent harness on long-running tasks. The poster explained that the harness needs to handle general research and other tasks that run longer than half a minute, explicitly excluding coding, and that they want the setup to perform better on that kind of workload. The harness already includes several relevant capabilities: the system prompt is rendered in real time and can be adjusted in response to almost any circumstance; an effective skills system is in place, allowing instructions to be written for any task with little effort, with the agent picking them up as needed; and the agent has access to web search and a Chromi tool. According to the thread, the discussion also covers real-time system prompt rendering, skills system design, and methods for reducing human intervention in complex long-running tasks. At this point the exchange remains at the stage of the poster publicly asking for advice on optimizing the harness, with no concrete improvements or conclusions yet offered in the discussion.