Can a coding agent build a working app overnight while you sleep? In May 2026, 42 Berlin ran a hackathon to find out.
what a coding agent is.
A coding agent is an AI program that writes code, runs the tests, reads the errors and fixes its own work in a loop, with nobody at the keyboard. The question was whether one could keep that up all night, alone.
building the builder.
Thursday was for building the builder. Teams spent the day designing their own autonomous coding agents: orchestration, tools, prompts, recovery logic, the whole stack.
8pm to 11am.
At 8pm a sealed engineering spec was released. From that point the agents were on their own. They had until 11am the next morning to read the spec, build the software it described, and leave behind enough logs and commit history to show their work.
The spec turned out to be a compiler for a small knitting language: read a knitting pattern, turn it into structured data, pass a test suite scored out of 150.
The twist: no paid API tokens. Local models and free-tier inference only. Public repos from participating teams show agents running open models such as Qwen3-Coder on their own machines. That moved the challenge away from calling the most powerful model and towards a harder question. How do you design a system that holds together overnight, unattended, on limited resources? Some did. Some didn't. That was the interesting part.
the winners.
Faris, Manoa, Jonas, Vincent, Hossam, Heléne, Isabel, Kabuniang and Nwe Thiri, across the prize-winning teams. Anastasia organised the hackathon, and Jan from Needle sat on the jury.
Needle builds an AI-powered search and agent development platform that connects tools such as Slack, Gmail and Confluence. It was founded in 2024, works from Berlin and San Francisco, and raised a $2.2 million seed round in 2025.
Faris, Manoa, Jonas and six others