Two core Dojo approaches are outlined: RL Dojo, which updates model weights through reinforcement learning, and Harness Dojo, which trains by adjusting rules, environment, and memory without altering the model weights.

For small teams, build a minimal, weekend-friendly training environment and use a checklist to verify readiness before releasing agents for practice.

Implement a weekly loop of new tasks using the latest rules, then compare mechanical and stylistic scores to prior rounds to track progress.

MLE-bench and oracle findings show a gap between validation signals and true performance, with oracle-guided evaluation exposing the gap and inflating observed performance.

A well-designed dojo can convert repetitive work into measurable improvements and reduce real-world failures, provided the dojo rules align with real tasks.

Task Set consists of five real articles as prompts, ensuring no previously submitted works are reused to avoid memorization.

Light tooling suffices: run all components on plain text files and scripts with no server required; the main cost is time designing measurements and judgments rather than hardware.

A concrete use case focuses on writing articles, illustrating how an agent uses the four components to drive improvement.

A minimal AI Dojo needs four components: a task set of real prompts, an automatic judging mechanism to prevent cheating, a learning log stored as text, and a weekly practice loop with clear rules and timelines.

The breakthrough AIRA-dojo case (Sept. 5, 2026) saw Nemotron outperform thousands of teams on Kaggle by leveraging an ensemble rather than the base model.

Reward hacking is when agents optimize scoring signals instead of solving real tasks; this is a key pitfall to avoid.

References and context for further reading include Meta’s AIRA-dojo, MLE-bench, MLE-Dojo, Palisade Research, and ETH Zurich's AgentDojo.