About
AI is moving from answering questions to doing work.
That shift is far harder than it looks. Models can already reason, search, and code — that part is largely solved. But the best work in any field has no answer key. It takes iterating within real constraints, deciding under uncertainty, and evolving across the edges of what's known.
Our mission
Navers Lab exists to study, rigorously and systematically, the limits of what AI can do on real work — and to help move those limits forward.
Where we're going
We believe that within five years, AI agents will be core collaborators in scientific research, engineering design, and professional decision-making.
- We've already seen agents become genuinely capable at software engineering — built a particular way: a harness that ships as a product and learns from real user trajectories, paired with large-scale reinforcement learning.
- We're now seeing agents trained for AutoResearch, growing steadily stronger at following long-horizon goals and learning from the feedback their environment gives back.
- We expect the same approach to reshape at least twenty more expert crafts in the years ahead.
What we do
We work on three fronts, each feeding the others:
- Benchmark
- Rigorous, real-world benchmarks that measure where agents actually stand on professional work, not toy tasks.
- Build
- The harnesses, agents, and training pipelines that move agents from operating tools to doing the work well.
- Break
- Stress agents to their breaking point, find where they fail, and push past it.