The Creators of Claude Code and OpenClaw don't Prompt Their Agents Anymore?!
The best AI engineers barely prompt their agents anymore. Sounds crazy I know! They build a loop that does the prompting for them and let it run for hours while they step away. The head of Claude Code at Anthropic put it bluntly: he doesn't prompt Claude, he has loops running that prompt Claude for him. The only problem is this gets expensive (and hallucinations compound) - you need a harness with observability that can run reliably not always using the expensive frontier models. So I built one! It's a TypeScript app that runs Pi (on Kimi) in a loop, stores every run in a database, and shows it all in a dashboard. And the loop isn't one agent on repeat, it's agents prompting agents: an orchestrator decides the next tasks and spins up workers to run them in parallel. Think of it like an advanced Ralph loop with more control! ~~~~~~~~~~~~~~~~~~~~~~~~~~ - Build and deploy your AI built applications easily with Retool - plus when you sign up now you get free app imports till July 1st and bonus AI credits on all paid plans! https://fandf.co/4uh4uDD - Neon, my favorite Postgres platform: https://get.neon.com/LqufgGN ~~~~~~~~~~~~~~~~~~~~~~~~~~ - The Dynamous Agentic Coding Course is now FULLY released - learn how to build reliable and repeatable systems for AI coding: https://dynamous.ai/agentic-coding-course - Agent Control Plane (my open-source Pi loop dashboard): https://github.com/coleam00/agent-control-plane ~~~~~~~~~~~~~~~~~~~~~~~~~~ 0:00 The Loop Engineering Buzzword 1:42 The Core Concept of Loops 5:52 Downsides and Token Costs 8:34 Deterministic Workflows with Archon 13:18 Orchestrating Parallel Coding Agents 17:29 My Pi Loop Engineering Dashboard 21:11 Deploying Control Systems to Production 23:58 Outro ~~~~~~~~~~~~~~~~~~~~~~~~~~ Join me as I push the limits of what is possible with AI. I'll be uploading videos weekly - at least every Wednesday at 7:00 PM CDT!
