Harness Engineering: What Actually Runs Your AI Agent

NVIDIA published a result in August that should change how you think about building agents. Their AVO system cleared all 183 levels of the ARC-AGI-3 public set with a perfect 100.00 efficiency score. The model underneath was Claude Opus 5, which scores around 30% on that same benchmark by itself. NVIDIA didn’t train anything new. They wrapped an existing model in their own agent architecture and got a 3x jump.

That architecture has a name. It’s called a harness: the agent loop, the tool registry, context management, sandboxing, permissions, error recovery. Everything sitting between the model and the actual work. It’s where most agent projects quietly fall apart, and it’s what this session is about.

Two weeks before the NVIDIA post, DeepSeek open-sourced theirs. DeepSeek Harness (dsh) is MIT licensed, built on the Cordis meta-framework, and organized around one idea: everything is a plugin. Models, tools, skills, sessions, sandboxes, loops, orchestration, even the UI. All of it swappable in config without touching the source. It collected tens of thousands of GitHub stars within days.

Meanwhile Claude Code, Codex, and Cursor are solving the same problems behind closed doors.

We’re also looking for speakers for upcoming events.

If you’re building with agents, running something in production, or sitting on a war story worth 30 minutes, add your name here:
[AI Agents Meetup Speaker Signup – Fill in form]. It doesn’t need to be as deep as this session. Practical beats polished, and half-finished experiments make better talks than finished ones.