February 9, 2026
From IDE to IDK
Eight More Months of Agents
Coder claims bots write 90% of his code; comments panic, clap back, and ask why learn now
TLDR: Author says AI agents now write most of his code, and he’s ditched modern tools for old-school Vi. The comments explode: warnings about corporate-controlled AI, clapbacks comparing bans to outlawing power tools, and a practical worry—why learn today if the frontier resets in months?
A veteran coder says AI “agents” have leaped from helpful sidekicks to full-on code-writing machines, bragging a frontier model now drafts nine-tenths of his work. He’s even ditched his IDE (fancy coding software) for old-school Vi, calling public tests “gamed” and cheaper models like “Sonnet” actively harmful for learning. Cue the internet meltdown.
One camp is alarmed: almostdeadguy warns of “massively negative repercussions” if a few giant companies control the tech, not apocalypse-level doom, but society-scale mess. Another camp claps back with construction-site snark: hasperdi says banning AI is like “outlawing power tools in carpentry,” sparking replies about safety rules, unions, and who gets paid when robots do the nailing. Pragmatists jump in with treadmill anxiety: dirkc asks why learn now if the best models change every eight months. Meanwhile, happytoexplain insists the divide isn’t complicated—everyone knows the stakes, they just disagree. HN’s resident archivist dang drops a link to last year’s agent debate, implying we’re in season two of the same show.
The memes? “From IDE to IDK,” “Vi turns 50, now with a robot intern,” and “benchmarks are WWE—fun, but scripted.” The drama? Who gets power from these power tools—and who gets left holding the hammer.
Key Points
- •The author reports a major shift in agent effectiveness driven by model improvements, while agent harnesses have not advanced much.
- •Coding capability rose from about 25% with Claude Code last year to about 90% with the latest Opus model this year, based on the author’s experience.
- •The author claims public LLM benchmarks are unreliable and emphasizes qualitative gains observed in practice.
- •They have moved away from IDEs, now using Vi and Neovim with minimal features (go‑to‑definition), noting Vi’s 50th anniversary in 2026.
- •The author warns that using non‑frontier models (e.g., Sonnet or local models) can teach incorrect lessons, though Mixtral made local models somewhat usable.