Proactive agents: the Context Graph proposal for enterprises

Today's enterprise agents share an uncomfortable limit: they wait. A salesperson has to ask about the status of a deal, an engineer opens the dashboard once there is already an incident, and a lawyer reviews a contract because someone flagged it. A new paper on arXiv, Context Graphs for Proactive Enterprise Agents, starts from that observation to defend a concrete thesis: the real productivity gain does not come from agents that answer better, but from agents that speak first.

The distinction is not rhetorical. The authors separate two modes of operation. The reactive mode dominates today in RAG and in agentic frameworks: the system waits for a human query and acts on it. The proactive mode reverses the order: the agent watches the state of the organization and surfaces actionable information before anyone asks for it. To support that leap they propose a central piece, the Context Graph.

What a Context Graph is

The Context Graph is a live relational structure that models the company's entities (contracts, incidents, sales opportunities), their relationships and, above all, their state transitions over time. It is not a static knowledge graph: its value lies in recording changes. On that base, the paper defines four components.

The first is a Delta Detection Engine that continuously monitors state changes in the graph. The second, a Proactivity Scorer that ranks candidate insights by urgency, relevance and fit with the recipient's profile (what they call persona fit). The third is a Surfacing Layer backed by an LLM that delivers ranked notifications with grounded explanations, that is, with the reasoning attached to the data. The authors formalize each block and derive a unified Proactivity Score function that combines those signals.

The implementation does not stop at the diagram. The work includes complete end to end Python code using NetworkX for the graph and the Anthropic Claude API for the generation layer. That decision matters for anyone wanting to reproduce it: NetworkX is a mature and accessible library, and grounding the language layer in Claude keeps the system within a stack many teams already have in place.

Why it matters and for whom

The evaluation rests on three generic enterprise use cases: contract lifecycle management, engineering incident response and sales pipeline hygiene. These are three domains where the cost of finding out late is high and measurable: a clause that lapses without renewal, an incident that escalates because no one saw the early signal, an opportunity that goes cold in the CRM. On that ground, a system that scores and prioritizes what deserves a notification has a clear fit.

It is worth reading the proposal with a cool head. The paper describes an architecture and an evaluation over generic cases, not a large scale production deployment with long term business metrics. The obvious risk of any proactive system is noise: an agent that alerts too much becomes a filter people learn to ignore, and there the Proactivity Scorer carries almost all the weight. The quality of the ranking, and the ability to calibrate the threshold per persona, will decide whether this saves time or consumes it.

For teams already working with Claude, the practical interest is twofold. On one hand, the pattern of state graph plus delta engine plus language layer is replicable with open tools and an API they already use. On the other, it fits the direction MCP servers and subagents are pointing toward: systems that not only answer in a chat, but observe sources and act when needed.

At ElephantPink we have spent time building integrations where the value is not in the model that answers, but in the circuit that decides when it is worth speaking. The Context Graph gives a name and a structure to that problem, and that is its most useful contribution, beyond the fact that the specific execution still has to prove itself outside the lab.