16h • General discussion
Coding tip for loops
from another user....
The default Clawdbot approach for coding is sub-agents. They work, but they share context, accumulate hallucinations, and eventually start confidently refactoring code that was fine. So we override it. Instead of sub-agents, I spawn Codex CLI directly — OpenAI's coding model in full-auto mode. Different model, different sandbox, way more reliable for implementation work. Then we wrap it in Ralph loops: a retry loop with a PRD checklist. Each iteration gets completely fresh context. The agent discovers project state from files and git history, not from memory of what it did last time. Says it's done? Loop validates against the checklist. Lies about completion? Restart. Stalls? Restart. Dies? Restart. Yesterday I ran three of these in parallel on separate git worktrees. 108 tasks across three branches in 4 hours. Each agent completely unaware of the others. The key insight: amnesia is a feature. Fresh eyes every iteration means no compounding hallucinations. The PRD is the source of truth, not the agent's memory.
what works for you ?
0
0 comments
Coding tip for loops
powered by
skool.com/openclaw-8860
Chat and learn together about OpenClaw, Autonomous AI agents, AGI and how to benefit
Suggested communities
Powered by