Two days of work six months ago, a few hours today: researcher benchmarks agentic coding's leap
avt_im · x · 2026-09-27
Researcher Alexander Terenin's new blog post measures how fast agentic coding has moved: his年初 CUDA hash table reimplementation took two days with human-in-the-loop AI; a similar-difficulty project now takes hours.
Case study: building a Rust PDF compressor from scratch to tame 1000+ accumulated ML papers. He had Claude reverse-engineer why ILovePDF works so well (yielding 15 techniques: image resizing, font merging, metadata removal), then adopted evaluation-driven development — assemble a 5,000-PDF test suite covering edge cases first, then fire off a single prompt. He went hiking for hours and returned to a working codebase competitive with ILovePDF on his thousand-paper benchmark, spending half a week's token budget.
His takeaway: tokens keep getting cheaper for a given usefulness level; if you haven't run such an experiment and updated your worldview, do it soon — your sense of what's easy, hard, or possible will change. Code is on GitHub and crates.io (cargo install compress-pdf).
More from coding & agent
- Google's Agent Substrate Deep Dive: 250 Stateful Agents on 8 Pods, Plus Egress Gaps — ghumare64 · 2026-09-27
- Hugging Face Engineer: AI-Generated PR Floods Are Pushing Open Source Into Hell — RisingSayak · 2026-09-27
- Devs rip Claude Code as least proactive coding agent, switching to Codex — TheMoonMidas · 2026-09-27
- Two-month deep-dive review of Factory's Droid coding agent, Pro vs Max tiers — Scobleizer · 2026-09-27
- Palantir's agent architecture docs decoded: six design decisions for a reliable agent harness — blaizedsouza · 2026-09-27
- open-slide 2.0 ships: an agent-first slide framework with visual editor and editable pptx export — CurieuxExplorer · 2026-09-27