Behind Navigator n2: How did Yutori make AI agents 10x cheaper?
Workshop #6 in the TwoSetAI Workshop series.
Building AI agents that operate your computer used to cost $10-30 per task at frontier scale. Yutori's Navigator n2 dropped that to $1.46, while topping state-of-the-art on 4 of 5 major CUA benchmarks. Same tasks, roughly a tenth of the compute.
In this 60-minute workshop, Skanda Vaidyanath, founding AI researcher at Yutori, walks through the recursive training loop the team built to ship n2. Yutori used CUAs to generate over 10,000 training tasks across hundreds of applications in two months, with agents stress-testing their own verifiers to surface reward hacks and edge cases. Then trained a 27B model that beat 500B+ competitors on real desktop workflows.
If you're building agents that touch the browser, the desktop, or any GUI, this session will change what you invest in next.
What you'll walk away with:
The recursive training loop pattern: how to use agents to generate agent training data (reusable for any agent stack, not just CUA)
How Navigator n2 interleaves GUIs, CLIs, tools, and code intelligently (and why GUI-only agents are unnecessarily constrained)
OPSD (on-policy self-distillation): the technique that pushed peak performance 4x faster than group-rollout RL
Real production numbers on cost-performance trade-offs at the 27B scale
Who this is for:
Engineers building browser or desktop agents (Playwright, Selenium, custom stacks)
Founders shipping agent products where cost per task matters
ML researchers working on RL for long-horizon tasks
Anyone deciding whether to build vs. use a CUA framework
Format: 60 minutes, live on YouTube. ~40 min technical deep-dive, ~15 min live Q&A.
About the series
Workshop #6 in the TwoSetAI Workshop series.
Session #1 (July 17) Why multi-agent systems fail
Session #2 (July 31) Inside AgensFlow
Session #3 (August 14) Token Budgets for Coding Agents
Session #4 (August 28) Self-Improving Agent Loops
Session #5 (September 18) Building A Personal Agent That Actually Works and Sounds Like You
Session #6 (September 25) Behind Navigator n2: The Recursive Training Loop That Made a 27B Model Beat GPT-5.6 and Claude
See all upcoming sessions: https://twosetai.com/lab/workshops/
About the instructor
Skanda Vaidyanath is a founding AI researcher at Yutori, where he works on reinforcement learning for LLM post-training on long-horizon tasks. He contributed to Navigator n2, Yutori's 27B parameter computer-use model that sits at the cost-performance frontier and beats larger models on 4 of 5 CUA benchmarks. Skanda published at NeurIPS 2022 (LISA: Learning Interpretable Skill Abstractions) and his research spans hierarchical RL, credit assignment, and learning from natural language feedback. Previously: MS in Computer Science (AI track) at Stanford, BE in Computer Science at BITS Pilani Hyderabad.
About the host
Angelina Yang runs TwoSetAI and hosts TwoSetAI and Decoding AI Builders, long-form interview platforms where she's interviewed 100+ AI founders shaping this era. fast.ai fellow. 2024 Anthropic Build Tournament winner. Curates the workshop series and connects the speakers with builders who care.
About Yutori
Yutori builds state-of-the-art AI computer-use agents. Co-founded in 2024 by ex-Meta AI leaders Devi Parikh (Co-CEO), Abhishek Das (Co-CEO), and Dhruv Batra, the company launched from stealth in March 2025 with a $15M seed round led by Radical Ventures, with angel checks from Fei-Fei Li, Jeff Dean, Elad Gil, and Sarah Guo. Their Navigator model line evolved from browser-only agents to Navigator n2, released August 2026 as their first full-desktop model. Their Co-CEO Devi Parikh is a previous guest on TwoSetAI. Her episode on browser agents is here: https://twosetai.com/watch/devi-parikh/
Hosted by TwoSetAI.
Past events replay
Workshop #1 with Nicole Königstein — Stop Your Agents From Spinning — walked through why multi-agent systems fail in production.
Workshop #3 (Aug 14): Token Budgets for Coding Agents
Workshop #4 (Aug 28): Self-Improving Agent Loops