A way to run an agent for a long time based on the local file system + git warehouse: Initialize the overall goal of the agent, task list/function list, and then repeatedly read the external task list and update the progress in a Harness loop:
1. Re-read the full content for each session, so that parallel work can be avoided to avoid repeated tasks; 2 Each context window first reads the progress progress git log to obtain the tasks that need to be executed this time. 3 Perform smoke testing. 4 Select a task and execute it. 5 Perform end-to-end testing. 6 Save the results to the external file system and git log. 7 Continue this cycle until all tasks are completed. This mechanism is very simple and practical, with a clear architecture. It maximizes the use of the existing file system and git mechanism. It is a very smart and efficient method.