Stop Prompting, Start Looping: Principles of Loop Engineering
summary
Describes a loop architecture that runs an agent headless under deterministic guards, commits each tick, uses cheap iteration gates, a reviewer LLM, and hidden heldout tests to verify real fixes; ships a runnable LoopKit framework and recommends git worktrees for parallel runs; measures reliability with pass k = 30% (3/10) as an example, and enforces dollar/iteration caps, evolution seeding, and extractable skills to improve success.
AI-written summary · about
chapters (20)
- 0:00Loop engineering intro
- 1:13Beyond one-to-one prompts
- 2:35Loop design questions
- 3:40Layered agent context
- 4:42Headless agent runs
- 5:45Safety guards in loops
- 6:49Local commits for progress
- 7:31Iteration gate testing
- 8:13Second-opinion reviews
- 9:16Held-out validation gates
- 10:41Termination and budget limits
- 11:44Human oversight before merge
- 12:27Scaling with work trees
- 13:51Batching for concurrency
- 14:57Loop reliability with pass K
- 16:42Evolving solutions iteratively
- 17:25Skill libraries for agents
- 18:10Core loop principles
- 20:01Looping vs prompting
- 21:06LoopKit framework
related talks
01h 48mRalph Loops: Build Dumb AI Loops That Ship
52mBuild a Prompt Learning Loop- 52mBuild a Prompt Learning Loop
39mThe Boringly Simple Loop Powering GenAI Apps
33mHumans on the loop, not in it: Taking agentic engineering end to end- 33mHumans on the loop, not in it: Taking agentic engineering end to end
26mThe Extract/Validate/Extrapolate Loop
34mThe Human-in-the-Loop is Tired