One Prompt vs Six: I Ran Two AI Builders in Parallel at a Hackathon | by TheProdSDE | Jul, 2026
The Setup: One Person, Two AI Builders, Running at the Same Time
This was a solo, in-person hackathon — PromptWars, hosted by Google for Developers — so there was no team, no division of labor, just me and a 3-hour clock. The brief: Destination Discovery & Cultural Experiences.
Build a GenAI-powered platform that helps travelers discover destinations and engage with local culture in meaningful ways. The solution must use Generative AI to recommend attractions, uncover hidden gems, generate immersive storytelling, promote heritage, suggest local events, and connect visitors with authentic cultural experiences.
No mock data. Everything real, everything connected.
I want to be upfront about what I actually did, because the honest version is more useful than a clean narrative: I didn’t run one approach, watch it fail, and then switch to another. I ran two things in parallel.
I gave the Hermes agent — using Nvidia’s Nemotron 3 Ultra 550B-A55B model for reasoning via an Nvidia API key — a single prompt: just the raw challenge text, pasted as-is, nothing more. Hermes started working on its own from that one prompt. While it was running, I switched over and started refining prompts through Claude and feeding the refined specs into Gemini IDE’s Antigravity, iterating there.
About 45 minutes in, after a few iterations, Antigravity showed me a working UI with services actually integrated and calling real APIs. Hermes, meanwhile, was still running — still reasoning, still working through the problem with Nemotron 3 Ultra underneath it.
I want to be honest that this isn’t an apples-to-apples comparison. Hermes got one short prompt and had to infer almost everything itself(past skills and my interaction learning). Antigravity got a much more detailed, iteratively-refined spec, developed through several rounds with Claude. That’s not a fair fight in either direction — it’s closer to two different questions: “what can an agent do with almost nothing” versus “what can a code-gen IDE do when it’s handed a real spec.”
How each agent handled 429s during code generation
This is worth being precise about, because the difference is meaningful.
Both agents hit 429 quota-exceeded errors during the code-writing phase — not in the final app, but while the AI was actively generating code and calling the Gemini API to do so.
Hermes hit these 429s during its own reasoning loop and recovered autonomously. It self-detected the rate limit, backed off, and kept working — all from the single original prompt, with no intervention from me. That’s genuinely autonomous agent behaviour.
Antigravity had a general error-detection and self-fixing loop, which worked well for most issues during code generation. But for 429 specifically — detecting rate-limit errors and implementing backoff in the generated code itself — that wasn’t automatic. I had to write and send a dedicated rate-limit prompt to get that behaviour explicitly. The general error loop was there; the 429-specific backoff needed its own spec.
This is a meaningful distinction: Hermes’s autonomous recovery from 429 during code generation versus Antigravity needing an explicit spec for it is the clearest evidence that the two approaches differ fundamentally in how much they need from you.
Repo submitted for Hermes / Nemotron 3 Ultra attempt: github.com/karangehlod/cultural-travel-companion
The Hermes + Nemotron 3 Ultra (550B-A55B) run, working from a single pasted prompt — no follow-up iteration, no refinement.
The build I actually submitted to the hackathon was the Claude-to-Antigravity one — TravelYarro — and that repo, promptwars-indore, is the final submission.
Live app: travelyarro.web.app
Submitted repo: github.com/karangehlod/promptwars-indore
Hermes repo: https://github.com/karangehlod/cultural-travel-companion
Stack: Vite + React 19 + TypeScript (strict) + Tailwind + Zustand + Zod + @google/generative-ai, client-only, deployed to Firebase Hosting.
