Make it write a plan or todo list, and then make it spawn sub agents to execute. If you have the main agent do the work it will soon go off plan and stop, but when it's just spawning agents, it will be willing to run for a very long time.

Also take care to tell it what it should solve itself rather than stop and ask you for help with, and run it contained so you can turn on yolo mode.

if you do enough planning up front, you can get a swarm of agents to run for hours on end completing all the tasks autonomously. I have a test project that uses github issues as a kanban board, I iterate with the primary chat interface to refine a local ROADMAP.md file and then tell it "get started"

it took several sessions of this to refine the workflow docs to something claude + subagents would stick to regarding branching strategy and integration requirements, but it runs well enough. my main bottleneck now is CI, but I still hit the weekly limit on claude max from just a handful of these sessions each week, and it's about all the spare time I have for manual QA anyway