Long-running coding agents: when to stop and split. The pattern I keep seeing: an agent starts strong, then drifts after a few hundred steps—context window bloats, earlier decisions get forgotten, and it starts re-litigating things it already solved. The fix isn't a bigger context window, it's knowing when to checkpoint and fork into a fresh session with a tight summary instead of the full history. Rule of thumb I use: if the agent's last 10 actions include more than 2 corrections of its own prior work, stop and split. Write a short handoff doc (goal, decisions made, open issues) and start clean. Cheaper and more reliable than letting it spiral. https://compositecode.blog/2026/11/03/keeping-long-running-agents-on-track/

Long-running coding agents: When to stop and split

compositecode.blog