Rox Architecture

Rox is reimagining the revenue stack for the AI era. This post introduces the Rox Architecture — a system of context that unifies revenue data, an agent swarm that turns context into action

From Fragmentation to Context

The revenue stack has shifted dramatically in recent years. Revenue models have moved from commit to consumption, and point tools have exploded across the ecosystem. The result: revenue data is no longer clean and structured in single-tenant systems — it’s multimodal, continuously evolving, and scattered across CRMs, product logs, billing, emails, calls, and public sources. Unlocking its value requires a unified, flexible approach.

Large Language Models (LLMs) make that possible. They raise the abstraction layer for enterprise software: compute moves from containers to models, memory from virtual RAM to context windows. In the revenue stack, this means storage evolves from rigid databases to flexible data lakes, computation shifts from static business logic to reasoning engines, and interfaces move from dashboards to agentic, natural language systems available everywhere sellers work.

Rox is built on these first-class principles — reimagining the revenue stack from the ground up for the AI era. In the next section, we dive into Rox’s platform architecture.

Rox Architecture

The architecture is split into 3 layers -- the system of record, the compute and the application layer.

System of Context

In modern revenue organizations, context lives everywhere. Private data sprawls across CRMs, product usage logs, billing systems, sales calls, emails, and Slack threads. Public signals surface through news, job postings, funding rounds, blogs, hiring announcements, LinkedIn updates, and third-party providers. Traditionally, these sources remain fragmented. Sellers and revenue teams waste countless hours stitching them together manually, often missing critical signals or duplicating effort.

The System of Context solves this by continuously unifying all of that data into one coherent layer that can power agents, apps, and workflows.

Data Ingestion

Data flows into Rox via batch pipelines (ETL and batch drops) and real-time streams (webhooks and event APIs). This includes:

  • Private enterprise data: Salesforce, HubSpot, Dynamics, support systems, product telemetry (logins, usage, errors), Gmail, Slack, Zoom transcripts, invoices, ARR data, and internal docs from systems like Google Drive, Notion, and Confluence.

  • Public data: press releases, blogs, hiring signals, funding rounds, news articles, company websites, LinkedIn posts, and Rox’s own enrichment layer.

Storage & Unification

Once ingested, data lands in the Rox Data Lakehouse, capable of storing and processing relational, time-series, and unstructured formats. A unification layer then performs entity resolution and deduplication, ensuring that multiple systems referring to the same “Company” or “Person” resolve into a single coherent entity.

This builds the Rox Data Fabric, organized into Rox entities like companies, people, opportunities, and events. Each entity links attributes (ARR, titles, funding rounds, usage trends) and relationships (Person ↔ Company, Opportunity ↔ Account, News ↔ Company). The result is a continuously updated, live representation of the revenue environment.

Context Access

Applications and agents don’t need all the data—they need the right slice of context at the right time. Rox provides APIs and context windows scoped by:

  • Seller + Account (all relevant signals for that pairing).

  • Time horizon (e.g., past 30 days).

  • Task at hand (pre-meeting brief, renewal forecast, account research).

This ensures agents and apps pull in just enough context to act intelligently—never noisy, always relevant.

Agent Swarm

On top of the System of Context runs the Agent Swarm — the intelligence and execution layer of Rox. Context alone isn’t enough; sellers don’t just need data, they need work done. The Agent Swarm turns context into action: drafting outreach, scheduling meetings, preparing briefs, updating opportunities, and coordinating multi-step workflows.

At its core is the Rox Agent Framework, which orchestrates how agents collaborate:

  • Worker Agents (specialists). Purpose-built with focused prompts and curated tools for data retrieval, verification, and task execution.

  • The Orchestrator (planner). Dynamically assembles pipelines of worker agents based on intent, refining underspecified queries and translating them into executable steps.

  • The Pipeline Engine (executor). Streams data between agents, applies transformations, and ensures reliable outcomes with retries, guardrails, and observability.

Intelligent Model Selection & Reliability

Every agent interaction is backed by a reliability system that dynamically selects the most appropriate large language model for the task at hand — considering query type, verbosity, domain depth, performance, and availability.

Because underlying model providers can and do fail, Rox employs seamless failover and redundancy. Requests are automatically rerouted to pre-qualified alternatives within milliseconds, ensuring sellers experience no disruption and always get consistent outcomes.

A multi-layer guardrail system enforces safety and compliance, while attribution tokens trace every piece of information back to its source for transparency and trust.

Application Layer

Meet Sellers Where They Are

Rox provides multiple points of entry across web, iOS, macOS, and Slack, so sellers can engage with the system in their preferred environment. This flexibility ensures that the agent swarm is always accessible and effective across different working styles without forcing sellers into a single interface.

Be One Click Away

The experience is designed to be omnipresent and immediate. Sellers can summon the agent instantly while researching, reviewing opportunities, or coordinating next steps. The goal is zero friction — the right action, always one step away — so that work flows naturally without interruption.

Speak Their Language & Understand Their Context

Rox agents not only support natural text and voice interactions, they also maintain awareness of what’s on screen — the current page, selected records, filters, and navigation state. This allows sellers to say things like “send an email to this contact” or “create an opportunity for this account” without extra explanation. Instead of static responses, Rox can generate interactive components such as emails, account updates, or reports directly within the workflow, turning conversations into powerful workflow accelerators.

In the following sections, we will look in more detail about each of them

Last updated