$100K to build a coding agent that runs on your own machine

Every coding agent most of us use today depends on a big model running in someone else's cloud. This challenge asks what it would take to get that same capability offline, on consumer hardware.

The task is to post-train Gemma 4, Google's open model, into a coding agent that can find its way around a real repository and draft fixes to real software engineering tasks on its own. Humans still handle architecture, logic, and review.

There are two tracks:

1. Developer Agent Competition ($65K). You post-train Gemma 4 into an agent that resolves real engineering tasks.

2. Paper Track ($35K). You submit new, unpublished research, or you write up how you approached the competition.

Total prizes are $100,000, and the entry deadline is November 25, 2026.

This lines up with what we keep saying in here. An agent is only as useful as the loop around it: a plan before it touches files, tests it has to pass, and a human at the gate. A local agent makes that loop cheaper to run, and you don't have to send your code to a third party.

A few questions for the group:

- Is anyone entering? I'd like to put a small team together.

- Who has done fine-tuning or reinforcement learning on an open model before?

- What hardware are you running locally?

Not a vendor brochure. Just a good problem.

— Keith

1 comment

$100K to build a coding agent that runs on your own machine