Tutti parlano di Loop Engineering... ma nessuno te lo spiega così
The video traces the evolution of AI interaction paradigms from Prompt Engineering through Context Engineering, Harness Engineering, to the newest Loop Engineering approach. Loop Engineering involves wrapping autonomous AI workflows in iterative loops that self-improve toward defined goals without requiring manual intervention between steps.
Summary
The transcript discusses the evolution of AI engineering methodologies over time. Prompt Engineering was the initial approach where developers wrote detailed system prompts to instruct AI models on their role and behavior, with performance depending on prompt quality. Context Engineering emerged as models gained access to tools and could read files, search the web, and call APIs through protocols like MCP, but faced the problem of context rot—performance degradation when token counts exceeded 200,000. Harness Engineering solved this by structuring tasks into subtasks, using persistent file-based memory (markdown files) to maintain state between sessions, and reinitializing the agent at each step rather than compacting context. Loop Engineering, promoted by Boris Cerni (Cloud Code) and Peter Steinberger (Open Claude), represents the next evolution by placing entire Harness Engineering workflows inside iterative loops that run autonomously based on triggers and verification conditions. The system cycles through execution and verification phases, writing results to memory files and triggering new loops based on events (GitHub issues, emails) or scheduled times. Key to Loop Engineering's success is defining measurable goals with termination conditions to prevent infinite loops. The speaker presents five verification levels: deterministic (boolean pass/fail), rule-based (numerical constraints), delayed truth (results verified later), LLM-as-judge (model self-evaluates), and human checkpoint (human review). The practical benefit is developers no longer manually prompt "improve this, improve that" but instead set a goal and let the loop run autonomously. The video concludes with a real example of Loop Engineering optimizing matrix multiplication code, achieving 320x speedup over 10 automated iterations without manual intervention.
Key Insights
- Boris Cerni and Peter Steinberger shifted their approach from writing prompts directly to designing loops that automatically generate prompts, fundamentally changing how developers interact with AI agents
- Context rot degrades AI model performance drastically above 200,000 tokens due to increased confusion from excessive input, requiring external management strategies
- Harness Engineering uses file-based persistence (markdown files) as an external memory system to avoid context compaction penalties, allowing agents to maintain state across sessions
- Loop Engineering creates nested loops—tool-calling loops within Harness loops within outer iteration loops—enabling autonomous cycles that eliminate the need for repeated manual prompting
- Five verification levels exist for loop goals: deterministic boolean conditions, numerical rule constraints, delayed truth outcomes, LLM self-evaluation, and human checkpoints, each suited to different problem types
Topics
Transcript
[0:00] And here we are once again with a new term that is going super viral called loop Engineering. Everyone is really talking about it. We went from prompt Engineering to Context Engineering to Arnest Engineering which lasted very little and now we have already entered this new era of Open Engineering. What is it? Is it all hype? Is it marketing? we'll find out in this video. This new, let's say, paradigm was [0:31] launched by these two characters here in the Tech world and I who are super famous. Let's talk about Boris Cerni, the creator of Cloud Code, who in a recent interview said these words where he says that he no longer writes the prompts on…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Simone Rizzo
DeepSeek ha appena reso TUTTI gli LLM più veloci
DeepSeek's new Spark technique uses semi-autoregressive speculative decoding to accelerate LLM inference by 51-400% without quality loss or model retraining. By combining a fast parallel draft model with an efficient verification process, Spark achieves higher token acceptance rates than competing methods like Eagle 3 and Flash, enabling faster inference on consumer hardware.
GLM 5.2 gira in locale quantizzandolo ad 1bit! #intelligenzaartificiale #aiagent
Researchers successfully ran the 744-billion parameter GLM 5.2 model locally on a Mac Studio M3 Ultra using dynamic quantization, compressing it from 810 GB to 223 GB. The 1-bit quantized version maintains 76.2% accuracy while being 86% smaller, and performs comparably to closed-source models like Claude Opus and GPT-5.5.