How I got Claude Code and Codex to pursue goals over weeks

Claude Code and Codex are great at a single task, then they stop. You give one a big task, it runs for a while, finishes what it can in that session and then you're back to deciding the next step yourself. I realized you can't plan a long-term goal in one sitting. The agent has to keep running and updating its plans as it observes your environment. That's what’s blocking other agent harnesses such as OpenClaw or Hermes from being able to achieve your long term goals. So I built SmithersBot to keep going. You send it a goal from Telegram, it turns that into a plan you approve, and it works through the plan task by task. When a plan is done, it proposes the next one and keeps pursuing the larger goal over days, weeks, or months. It keeps running until an Observation Point; a time when either human judgement is needed to guide the future work or time needs to pass to observe the results of the agent’s work. **How I got my agent to keep going without giving up control:** * It asks me the key decisions up front and recommends an answer, so it isn't hallucinating what I meant. * Each task runs in a fresh worker, so a long session doesn't degrade. * It git checkpoints before every task, so a bad step can be ralphed or rolled back. * Build and test checks run outside the worker, so it can't just tell me it passed when it didn't. * If a task gets blocked, it keeps working on the ones that aren't waiting on it. * Every action and tool call written to disk, so the agents and I can see exactly what happened. I've been using it to improve itself and the next goal it will pursue is building and operating its own company. It's open source and free. What long-term goal would you want your agent to chase over weeks?