Ranked Coding, With the AI Agents Allowed In

Rochester, N.Y. Coding interviews still test whether you can reverse a linked list alone. Real work looks different now: you and an agent like Claude Code or Codex shipping something together. BeatCode is my attempt at measuring that skill. You pick a challenge in Scripting, Backend, Frontend or CyberSec, build it, and get rated.

There are two ways to submit. You can hand over a GitHub repo, and the server spins up a fresh E2B sandbox that lives for 5 minutes, clones it and runs the problem's trusted grading command. Or you can code right in the browser: a 30-minute sandbox runs VS Code (code-server), embedded in the page, and gets graded in place.

Grading is deterministic. The judge scores the sandbox output against expected markers and reports a 0 to 100 score with per-test results. Some problems test themselves, some use a trusted one-line verifier, and some ship a full grader script.

Scores feed a custom Elo system. Problems are rated 1000, 1400 or 1800 by difficulty, K is 32, and partial credit kicks in between 40 and 79 percent. XP and levels sit on top of that.

Version 2 is in progress. Each problem gets its own sandbox template with Claude Code and Codex preinstalled, and users can bring their own API keys, encrypted with AES-256-GCM and injected only into their own sandbox. The new graders go further: the React calculator one drives headless Chromium with Playwright, and the Node API one starts your server and fires HTTP requests at it. Failed sandboxes stay alive so you can fix and resubmit, and a GitHub Action rebuilds the sandbox snapshot on every push.