OVER DEZE AFLEVERING
What do coding agents need from engineers to be trusted? AI consultant Fausto Albers builds production software with agents and doesn't believe anyone still reads the code line by line. Here is what he relies on instead.
We cover:
- Why newer models are "eating the harness", and the shift from context engineering to loop engineering
- The state a coding agent should be in before it runs for hours, and the gates its output has to pass
- Observability with open-source MLflow when you can't read every trace
- A shared ontology and a three-layer memory stack: a Karpathy-style Markdown wiki in Obsidian, graph memory and a local vector database in Docker
- Disposable vs durable software, and why software built by intent threatens traditional SaaS
- Why AI makes attacks cheaper than defense
This episode is for software engineers and tech leads who want to give coding agents real autonomy without giving up responsibility for what gets built.
Timestamps:
00:00:00 - Intro
00:00:49 - Why Newer Models Keep Eating the Harness
00:02:26 - Will SaaS Survive the Next Ten Years?
00:03:39 - From Sociology to Cocktail Bartender to Restaurant Owner
00:07:19 - Why Human Preference Can't Be Reduced to Data
00:10:10 - What I Learned From a Failed Restaurant Chatbot
00:14:16 - Learning Faster With AI and Who Gets Left Behind
00:18:38 - What State Should Your Coding Agent Be In?
00:22:09 - Building a Shared Understanding Before Kicking Off a Loop
00:26:58 - Why Nobody Really Reads AI Code Line by Line
00:31:34 - The Real Reason Engineers Push Back on AI
00:33:48 - Sloppy, Disposable or Durable: How to Judge Code
00:35:18 - A Three-Layer Memory System for Coding Agents
00:39:44 - Why Domain Experts Matter When Execution Gets Cheap
00:42:23 - How AI Makes Attacks Cheaper Than Defense
Guest: Fausto Albers, AI consultant at Wonder Why
https://www.linkedin.com/in/stepintoliquid
