Architecture / Docs / Agent Swarm™ Layer
Human Agent Interaction
Over the last two years, we’ve seen AI adoption accelerate fastest where assistants can produce something and do something. This is most obvious in Software Engineering and Customer Support, but for Sales, there isn't a clear system of action.
Our Vision
At Rox, we want to pioneer systems that execute across the entire sales cycle. Delivering on this requires us to rethink the sales cycle itself and identify the moments where agentic capabilities create the most leverage. From there, we can imagine what it feels like to command not just one agent, but an entire coordinated Agent Swarm layer.
When designing Command, we grounded our work in three principles:
Trust & Clarity
Human in the Loop
Ease of Use
Trust & Clarity
We made sure that Command provides full visibility into its work, sources consulted, data pulled, actions queued, and results, so you always know what it’s doing and why.
Fine-grain Reasoning
When Command performs deeper work, like digging through notes, transcripts, emails, it streams its current step ("Searching over meetings with Ramp", "Enriching contact information", "Drafting a follow-up email", etc.) so you’re never guessing about progress.
Screen-Aware Context
When users interact with Rox, the agent captures screen context as part of the request: current page state, visible data elements, active filters, selected items, and navigation history. With this context, the agent can correctly interpret on-screen references, relationships between displayed data, and the user’s current workflow state.
Human in the Loop
At Rox, we believe in supercharging sales teams, not replacing them. Sales is a deeply personal, relationship-driven discipline, where human judgment makes all the difference. In practice, this means most agent actions require user confirmation at some point in the workflow. The agent should suggest and be proactive, but never overreach or make irreversible decisions.
Dynamic Action Components
Our vision goes beyond text generation. We designed Command to generate specialized, interactive components that align with Rox’s business capabilities. These components transform conversations into workflow accelerators that directly integrate with the platform.
We maintain a library of domain-specific components, covering opportunities, emails, research summaries, outreach campaigns, account profiles, and more. When responding, Command selects and configures the right component for the context, ensuring users receive actionable interfaces, not static text.
Ease of Use
Command is designed as a simpler abstraction over Rox’s core offerings. Instead of navigating a complex UI, users interact with Command through natural language and a conversational interface.
Workflow Suggestions
The agent UI framework is built around real user workflows within Rox. By understanding how users move between interface elements, Command anticipates needs and provides proactive assistance aligned with natural patterns.
For example, when viewing an account profile, the agent infers the likely next steps (creating opportunities, scheduling outreach, generating reports) and can interpret ambiguous requests based on that context. This removes the need for users to spell out every detail.
What's Next?
This was our first pass at building a different type of software at Rox. One that is simple but powerful behind the scenes, and an interface that reflects that. As we create more agentic experiences in other parts of the product, we'll be thinking about other ideas like proactivity, adaptability, and memory.