The paper proposes the Context Graph as a foundation for building proactive agents in the enterprise. The authors argue that while retrieval-augmented generation and agentic frameworks have advanced enterprise AI, they remain largely reactive-waiting for user queries instead of proactively surfacing actionable information.
RAG and agentic systems handle queries effectively, but they do not inherently improve worker productivity unless they can act proactively. Without mechanisms to continuously monitor and interpret organizational state, important signals are missed or delayed.
The Context Graph is a live relational data structure that models enterprise entities, their relationships, and transitions of state across time. By maintaining this evolving graph, systems can track dependencies between items such as projects, assets, and tickets, and observe how changes propagate through the organization.
- Delta Detection Engine: continuously monitors the graph and detects state changes that may deserve attention.
- Proactivity Scorer: ranks candidate insights by urgency, relevance, and persona-fit to prioritize what should be surfaced.
- Surfacing Layer: delivers ranked notifications with grounded explanations using a language model to generate clear, contextualized messages.
Practical deployment requires keeping the graph current, implementing continuous detection pipelines, and defining notification policies to prevent alert fatigue. Equally important are metrics for proactivity and methods to validate the grounding and usefulness of generated explanations.
The Context Graph architecture links a live relational model with continuous delta detection, a proactivity ranking mechanism, and a surfacing layer that produces grounded notifications. This combination targets the transition from reactive to proactive enterprise agents, enabling workers to receive timely, relevant actions before they explicitly request help.
1. What is a Context Graph? A Context Graph is a live relational structure representing enterprise entities, their relationships, and state transitions over time.
2. What does the Delta Detection Engine do? The Delta Detection Engine continuously monitors the graph and identifies meaningful state changes.
3. How does the Proactivity Scorer work? The Proactivity Scorer ranks candidate insights by urgency, relevance, and persona-fit to prioritize surfacing.
4. What challenges arise when deploying this architecture? Challenges include maintaining graph freshness, preventing notification overload, and validating generated explanations.
5. Does the Surfacing Layer explain notifications? Yes, the Surfacing Layer uses a language model to produce grounded explanations accompanying ranked notifications.
Source: https://arxiv.org/abs/2607.07721