Part of Hack Summit 2026. See the full summit: https://magicball.dev/hack-summit-2026
Turn your Claude Code history into an RL gym
Your own benchmark and your first RL experiment, built from how you actually work.
You've probably spent hundreds of hours working with Claude Code. All those sessions contain real tasks, decisions, mistakes and corrections.
What if you could turn that history into a benchmark for other models, and then use it to train your own?
In this Hack Hour, we'll try to recover tasks from your Claude Code history, build verifiable tests around them and benchmark open models on the kind of work you actually do.
Then we'll try something fun. Can we make a small 4B model better at your tasks using RL?
We'll run a short training experiment with Qwen3.5-4B and see how far it gets.
What we'll do
-
Turn your sessions into tasks. Extract real coding tasks from your Claude Code history and check which ones can be reproduced and verified.
-
Build your RL gym. Convert valid tasks into runnable Harbor environments with tests to check whether an agent actually solved them.
-
Benchmark models. Run open models against your tasks and see which ones perform best on your actual work.
-
Try RL. Use those tests as rewards to train a small Qwen model and compare it against the original.
What you'll leave with
-
Your own benchmark, built from your coding history
-
A leaderboard of models on your tasks
-
A hands-on RL experiment and the open-source tools to continue experimenting
Before you arrive
-
Modal: Create a free account at modal.com. Modal gives you $30 in free compute credits every month, which should be enough for the small-model experiments we'll run during the Hack Hour, including benchmarking and a short RL run. We'll share setup instructions before the event.
-
Laptop: Bring your laptop with Claude Code installed and some existing session history. We'll have sample tasks available in case your sessions can't be converted into runnable environments.
-
Prerequisites: Basic Python knowledge. No prior RL experience needed.
Format: 10 min demo · 40 min hands-on building · 10 min results and discussion
When: Saturday, 12 December 2026
Where: HSR Layout, Bengaluru. Exact venue shared after approval.
Seats: 30, approval required.
After the session
You'll keep your tasks and benchmark, plus access to office hours with us if you want to try this on your team's production agents.
About Hack Summit 2026
Magicball's Hack Summit is on Sat 12 Dec in HSR Layout, Bengaluru: 20 Hack Hours, 50+ booths and 1,000+ AI engineers in one building. Approval for this Hack Hour includes entry to Hack Summit 2026. Each Hack Hour has its own approval, so request a seat for every Hack Hour you want to join.
Hack Summit 2026 is in partnership with Grayscale Ventures and Testsigma.
See the full summit: https://magicball.dev/hack-summit-2026
More Hack Hours
Explore every Hack Hour: https://www.magicball.dev/hack-summit-2026#hack-hours