I Asked 3 AI Agents to "Make It Better" Until They Refused. Only One Did.
I tested Claude Opus 4.6, Gemini 3.1 Pro, and Codex GPT-5.4 to see if they'd push back when asked to keep improving already-finished code. The results split three ways.
Py4_ - A lens into the entropy of being
Personal blog about software engineering, life, and deep thoughts. Writing from Tehran to Canada.
Latest posts
The Priesthood of System Design
People say LLMs can code but can't do system design. I disagree. System design is more pattern-matching than people admit — and if you give an LLM enough context, it can outdesign 99% of engineers.
I Miss Coding
Coding agents have transformed my work at Google. But even if I keep my job, I don't enjoy this new role. I used to love figuring out the 'how.' Now the agent does it, and doing it myself is just inefficient.
Browser Sessions Are All-or-Nothing
Web sessions assume a human is in control. But when AI agents take over your browser, you're either logged in with full permissions or not. We need selective authorization—read-only modes, proof-of-human-intent, and agent-aware backends.