NVIDIA × Nebius Hack 2026

An agent-first discovery and delivery workspace for the
Coding and Agentic Engineering track.

Hackathon · Requirements · Research · Validate an idea · Research with ChatGPT · Ideas · Agent guide

Important

This project is in discovery. The user problem, product thesis, stack, and architecture have not been selected. Coding agents and harnesses are the exploration area—not a predetermined product.

What this repository is

This repository is the team's shared operating system for moving from evidence to a focused hackathon product. It keeps official constraints, external research, product hypotheses, accepted decisions, implementation, and evaluation artifacts separate so that humans and coding agents can collaborate without turning assumptions into facts.

The current investigation covers coding agents, harnesses, context engineering, agent runtimes, sandboxed execution, verification, state management, extensibility, and developer tooling.

Project signal	Current state
Track	Coding and Agentic Engineering
Phase	Discovery and problem selection
Product	Undecided
Architecture and stack	Undecided
Implementation	Intentionally not started
Submission deadline	October 30, 2026 at 10:00 AM PDT

Repository map

Path	Source of truth for
AGENTS.md	Repository-wide routing and collaboration rules
docs/hackathon.md	Official requirements, dates, judging criteria, and submission checklist
docs/product.md	The accepted product definition once the team chooses one
docs/architecture.md	Accepted architecture and its still-open questions
docs/evaluation.md	Evaluation policy, comparison discipline, and evidence requirements
docs/validate-idea-with-chatgpt.md	Fast ChatGPT gate against current hackathon rules and idea criteria
docs/research-with-chatgpt.md	Public entry point for researching an idea from a regular ChatGPT conversation
docs/decisions/	Architecture Decision Records for material choices
research/	Sourced external evidence and technical analysis
ideas/	Candidate problems and product hypotheses
app/	Future implementation, tests, scripts, and evaluations

Research baseline

The initial research package is ready to support product discovery:

• landscape.md compares Claude Code, Codex, DeepSeek Harness, OpenCode, OpenHands, mini-SWE-agent, SWE-agent, and Aider.
• deepseek-harness.md is the deeper architecture case study.
• architecture-patterns.md extracts cross-system harness patterns and their trade-offs.
• sandboxes.md records verified Nebius Token Factory Sandboxes capabilities and open questions.
• harness-primer.md provides a vendor-neutral shared vocabulary.

Research is evidence, not a decision. Any material product or architecture choice must be accepted explicitly and recorded in the canonical docs or an ADR.

How we work

1. Read AGENTS.md before making changes.
2. Check docs/hackathon.md before proposing anything that affects eligibility or submission.
3. Add sourced findings to research/; add unvalidated product hypotheses to ideas/.
4. Promote a direction into docs/product.md only after the team accepts it.
5. Record material technical choices in docs/decisions/ before implementation depends on them.
6. Keep empirical runs under app/evals/ and commit only curated, reproducible reports.

The working order is:

problem → user → insight → solution → architecture → technology → evaluation

Hackathon guardrails

The final submission must include working software that uses at least one NVIDIA open-source model and makes a runtime inference call through Token Factory or runs on eligible Nebius AI Cloud compute. It also needs a public source repository, setup instructions, a license, a hosted demo or test build, and an English demo video under three minutes.

See the team's operational requirements digest and the official rules before relying on this summary.

License

No license has been selected yet. Choosing and adding one is a required pre-submission decision.