A team that never touched the keyboard once won the hackathon. An AI agent wrote 100,000 lines of code overnight, 70,000 of which were test code. The developers just handed over the blueprint and went home. This actually happened in the middle of Seoul.
What Is This?
Ralphton is a hackathon built on the concept of "humans go home, AI codes." It was co-hosted by the developer community Team Attention and Kakao Ventures, with sponsorship from global big tech player OpenAI. It ran from February 28 to March 1, 2026 in Seongbuk-gu, Seoul, and OpenAI Singapore developers showed up in person.
Traditional hackathons have participants coding all night and then judging the results. Ralphton is completely different. Participants submit only their ideas and design documents (PRDs, specs), and the actual coding is done by AI agents. Once they complete the harness engineering on day one — basically designing the environment and constraints so the AI can work correctly — the AI agent autonomously codes overnight, and judges evaluate the results the next morning.
Nine teams made it through the qualifiers to the finals. Startup developers and founders participated, including Hwang Sunghyun (BankSalad co-founder, currently tech lead at TheVentures), who built an LLM-based puzzle web app. Kim Wooyoung, who runs an AI agent startup, developed a service that analyzes sheet music PDFs and automatically transposes keys.
What Changes?
The winning team’s case shows it most intuitively. They built a bot that automates household chores based on fixed-camera footage, but the key is in the process.
From start to finish, the participants never touched the keyboard once. Instead, before any development, they had the AI agents go through 133 rounds of Socratic reasoning with each other, bringing the ambiguity index of the design down to 0.05. The AI agents would ask each other things like "Does this requirement mean A or B?" and caught design gaps that humans might have missed on their own.
It’s also impressive that 70,000 of the 100,000 lines were test code. The AI didn’t take the "make it work first, test later" approach — it designed verifiable code from the start. The test code also preemptively blocked hallucinations that could arise from using multiple AI agents.
| Traditional Hackathon | Ralphton | |
|---|---|---|
| Developer Role | Write code directly | Blueprint + harness engineering |
| Who Codes | Humans | AI agents |
| Who Stays Up All Night | Developers | AI (developers go home) |
| Quality Assurance | Code review | 133 rounds of AI Q&A + 70% test code |
| Core Competency | Coding skills | Design ability + AI environment design |
There were failure cases too. One team hastily expanded specs mid-way and ran three work tracks in parallel, resulting in a deployment crash and an infinite loop. They ended up having to manually fix the code. It showed that giving AI agents too much autonomy without clear design can actually backfire.
The third-place team showed yet another approach. They focused on cost optimization, switching from expensive models to cheaper ones while maintaining the same performance metrics. It’s evidence that competitiveness in the AI agent era isn’t just about "using the best model" but also about "achieving the same results more efficiently."
In the past, there were clear limits to how a solo developer could scale a business, but through this Ralphton, we confirmed that barrier has completely fallen.
— Jang Dongwook, Director at Kakao Ventures
Team Attention developer Jeong Gubong also said, "Beyond using AI as a tool, how deeply you internalize AI agent utilization will become a company’s true moat."
The Essentials: How to Get Started
What Ralphton ultimately demonstrated is "harness engineering" in practice. Instead of writing code yourself, you design the environment for AI agents to work properly. Here are steps you can start with right now.
- Start by writing a precise PRD (Product Requirements Document)
The winning team’s secret wasn’t "not touching the keyboard" — it was "writing extremely precise design documents." Before handing things off to AI agents, eliminate as much ambiguity as possible from your requirements. If the question "Couldn’t this be interpreted another way?" comes up, your document isn’t ready yet. - Design verification loops between AI agents
There’s a reason the winning team had 133 rounds of Q&A. When one agent writes code, another agent should verify "Does this match the spec?" Set up this kind of loop. Start by putting verification rules in your Claude Code CLAUDE.md or connecting automated tests to your CI/CD pipeline. - Have AI write test code first
Flip the order from "code first, tests later" to "tests first, code after." It’s no coincidence that 70% of the winning team’s code was tests. You’re giving the AI agent criteria to verify its own output first. - Minimize spec changes, be cautious with parallelism
This is the lesson from the team that failed. Expanding specs mid-stream and running reckless parallel executions confuse AI agents. It’s better to push one clear work track to the finish line. - Accumulate harness engineering as a team asset
CLAUDE.md, MCP server configs, verification scripts — commit these to the repo, not just your personal environment. The compound effect only kicks in when the next project and the next team member can use the same harness.
What Is Harness Engineering?
It’s evolving from prompt engineering → context engineering → harness engineering.
Prompt = "What should I ask?"
Context = "What should I show?"
Harness = "How do I design the entire environment (scaffolding, constraints, verification loops) where AI operates?"
Ralphton is a case that proved harness engineering in the extreme environment of a hackathon.
