Skip to content

Loop Engineering

Lodri Péter edited this page Jun 25, 2026 · 1 revision

Loop Engineering

This project follows the Loop Engineering paradigm.

The ecosystem

Piece What Link
Addy Osmani's essay The canonical text — 5 building blocks + memory addyosmani.com
Cobus Greyling's reference impl npm tools (loop-audit, loop-init, loop-cost) github.com/cobusgreyling
LangChain: The Art of Loop Engineering 4 stacked loops (Agent → Verification → Event-Driven → Hill Climbing) langchain.com
LoopKit Python-native starter kit, Telegram bot, council github.com/peterlodri-sec

The four-tier loop stack (LangChain)

Tier Loop LoopKit primitive Role in kompress
L1 Agent Loop.run() training/eval iteration
L2 Verification evals/heretic.py adversarial grader
L3 Event-driven Telegram bot triggers on checkpoint events
L4 Hill-climbing LLM Council + trace rewriting decides SHIP/RETRAIN/PIVOT

The correctable loop invariant

If the evaluation metric doesn't improve in 3 iterations, stop and pivot. Applied to the self-labeling chain:

  • v3→v4: large jump (0.942→0.967) — continue
  • v4→v5: slight regression (0.967→0.961) — stop

Key quotes

"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." — Peter Steinberger

"I don't prompt Claude anymore. I have loops running that prompt Claude." — Boris Cherny (Anthropic)

Clone this wiki locally