DEV Community

Cover image for I ran two AI builders in parallel at a solo hackathon — here's what the 429 errors revealed
TheProdSDE
TheProdSDE

Posted on • Originally published at Medium

I ran two AI builders in parallel at a solo hackathon — here's what the 429 errors revealed

#ai #hackathon #typescript #promptengineering

At a 3-hour Google for Developers hackathon, I ran two AI builders simultaneously on the same brief:

  • Hermes + Nvidia Nemotron 3 Ultra 550B — one raw prompt, fully autonomous
  • Claude → Gemini Antigravity — iterative spec-first workflow

The most revealing moment wasn't about speed or features. It was how each one handled 429 rate-limit errors during code generation.

// The 3-layer strategy I ended up having to write explicitly for Antigravity:
const RETRY_DELAYS = [2000, 5000, 10000];

async function retryWithBackoff<T>(fn: () => Promise<T>): Promise<T> {
  for (let attempt = 0; attempt <= RETRY_DELAYS.length; attempt++) {
    try {
      return await fn();
    } catch (err: any) {
      const is429 = err?.status === 429;
      const hasRetry = attempt < RETRY_DELAYS.length;
      if (is429 && hasRetry) {
        await new Promise(r => setTimeout(r, RETRY_DELAYS[attempt]));
      } else { throw err; }
    }
  }
  throw new Error('unreachable');
}
Enter fullscreen mode Exit fullscreen mode

Hermes recovered from this autonomously. Antigravity needed the above as an explicit spec.

I scored 93%. Lost the remaining 7% to a single missing prompt — one I assumed I didn't need to write.

Full case study with every exact prompt I used (word for word):
👉 [Read on Medium → https://medium.com/gitconnected/i-ran-two-ai-builders-in-parallel-for-a-solo-hackathon-heres-what-actually-happened-ecb35bb7bafd?sharedUserId=theprodsde]

Includes: architecture diagrams, the 8-step build order, state persistence code, Zod schema tests, and the one prompt I wish I'd written.

Have you run multiple AI builders in parallel? What broke first?

Top comments (0)