Recently, I have been running long-horizon agent jobs — letting an agent keep working for a long time instead of babysitting every step.
The idea comes from the ralph loop: set things up once, then let the agent loop through the work until it is done. Before starting, I prepared three small files: prd.md for the product goal, agents.md for the working rules, and tickets.md for the task list. Then I let the agent run.
I tried three different approaches to do the work.
The first was Codex. I asked it to go through my tickets one by one and record the state and a review for each ticket. The work was fine, but the token consumption was not. I burned almost two complete weekly seat quotas very fast, and still got through only half of my tickets.
The second was DeepSeek with the PI agent harness. The problem there was different: the third-party DeepSeek API tries to overthink too much and keeps outputting complete code for a very long time. And I cannot monitor what is happening under the hood.
The third option, which is now my standard procedure, is to just run it with OpenClaw. I let it continue with my ticket list and go all the way to the end. It uses my common simple-code skill, which contains just an implementation sub-agent and a falsification agent. There is no fence set up — I just let the agents go through the tickets while keeping agents.md in mind. It really reports back to me when it finishes each ticket.
That was the most satisfying solution for running a long-horizon job. I can monitor it with my IM application, and I can steer it with slash-command tools. Running the agent loop becomes easy that way.
AI, OpenClaw, agent, long-horizon — Aug 21, 2026