-
-
Notifications
You must be signed in to change notification settings - Fork 0
Loop Engineering
Lodri Péter edited this page Jun 25, 2026
·
1 revision
This project follows the Loop Engineering paradigm.
| 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 |
| 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 |
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
"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)