Agentic workflow and closing the loop

The creator of OpenClaw: "I ship code I don't read"

Topic 5 of 8

Published Jan 28, 2026 · Clip 38:21–1:13:42 (35:21)

Peter details his agentic engineering workflow: being hooked on Claude Code like a slot machine, the 'black eye club' of friends up at 5 a.m., and discovering he can now build anything without mastering each stack. He explains feeling architectural friction through prompting, running 5–10 parallel agents, preferring OpenAI Codex (with GPT-5.x) for complex codebases because it reads extensively before writing, versus Claude Code's speed with more steering. Key principles: closing the loop so agents can verify their own work, letting agents write tests and documentation, designing codebases for agents, dismissing elaborate orchestration layers as waterfall, and under-prompting deliberately to spark new ideas.