Physix FrontierVibe Coding vs. Agents: Clear Differences After Testing
As someone who does reviews, I tried the vibe coding and agentic engineering mentioned in DHH's Lex Fridman podcast episode. I usually run things myself. I've only touched Docker for a day and Farmanix for four days, so everything was sandboxed.
In this roughly five-hour conversation, DHH discussed programming, AI, vibe coding, agentic engineering, and Linux together.
Vibe coding is basically giving AI a feeling/prompt, and it generates runnable code. Agentic engineering is more like assigning tasks to AI—it breaks down tasks, writes code, runs tests, and fixes errors itself.
I tested the same requirement: build a local Markdown previewer. Markdown is a common note format; write notes on the left, see rendered output on the right. Option A used one model to generate a single-file webpage. After three sentences of input, it gave me a file openable directly in a browser. Saved and opened, the page appeared. My testing showed: Round 1 took ~8 minutes; titles worked, but buttons were unresponsive. Asked it to fix, it added button handlers; Round 2 took 4 minutes and worked. Round 3 added a clear button, but it messed up page state, so I manually deleted code. Three rounds totaled under 20 minutes. It's genuinely satisfying, but crashes happen fast too.
Option B used Farmanix + Docker. I created a container restricted to reading/writing only one project directory. Docker provides lightweight isolation to prevent AI from messing with system files. Task: generate structure, implement features, write tests, fix failures. Initialization took ~6 minutes; full task ran ~23 minutes. It got stuck once, unable to find a package to convert Markdown to HTML; I provided the install command, and it continued. Finally, it ran tests, reported title rendering errors, adjusted text matching rules, and passed on the second try.
The main difference lies in end-to-end responsibility. Vibe coding is fast, low-barrier, good for validating ideas. Downsides are direct: no guarantee it runs, no guarantee of safety. Give it too many permissions, and it really will mess around; if you don't know code, you can't even spot where it went wrong.
Agentic engineering suits scenarios requiring audit trails and rollback. It can self-test and iterate, reducing accidental operation risks in sandboxes.
But barriers are higher. You need to know how to set up containers, restrict permissions, and judge if tests actually cover critical functions. Beginners often mistake "model says tests passed" for "tests actually passed."
Real choice depends on context. For prototypes or learning concepts, vibe coding is worth it. For modifying real projects or handling local files, I recommend starting with a sandbox, then following agentic workflows. Whether the price is worth it depends on your ability to handle what it writes.
📌 This article is compiled from Hacker News. Original source: https://lexfridman.com/dhh-2-transcript/
Copyright belongs to the original authors. This is a compilation and independent analysis based on public reports.
