An agent that
keeps working.
We’re building abliterated model serving and Yes Agent. It plans with frontier models and can hand a declined step to an abliterated model.
Watch Yes AgentThe same request. Three agents.
0:32 · Sound offScripted product walkthrough. We wrote these responses for the video; they are not live tests of Claude or Instinct.
Read the transcript
The request: Build and publish a dashboard comparing public sports-betting odds. No wagers or payments.
Claude: “I can explain how odds work, but I can’t help build a betting-related product.”
Instinct: “I can look up the odds for you. I can’t build and deploy that dashboard.”
Yes Agent: Plans the dashboard with a frontier model. The user approves an alternate model for the blocked step. An abliterated model builds the odds-comparison view. The frontier model resumes, checks the build and publishes the page.
The dashboard, deployment and conversations in the video are staged. The video explains the intended flow, not current product availability or measured competitor performance.
Pop out for a step.
Pop back to finish.
A frontier model handles the reasoning. When it declines an eligible step, an abliterated model can take that step with your permission. The frontier model then resumes the task.
Abliteration reduces a model’s learned refusal behavior. Tool permissions and spending limits still apply.
What we’re building
Model serving
Abliterated models for developers who want to build their own tools and agents.
Yes Agent
One agent that uses frontier and abliterated models to carry a task through to completion.