An arXiv preprint proposes a Context Graph that watches enterprise data; the architecture scores each change by urgency and uses Claude to draft a grounded alert before anyone opens a chat.
Most enterprise AI agents wait for the prompt. A new arXiv preprint from researchers working in the Anthropic Claude API ecosystem argues the next productivity gain comes from agents that interrupt, not agents that answer, with a graph-based architecture that watches business state and pushes ranked, grounded alerts before anyone opens a chat.
"Context Graphs for Proactive Enterprise Agents" is the architecture's name and its central claim. The paper does not treat proactive agents as marketing shorthand. It defines them as a pipeline that runs three jobs in sequence: see what changed, decide what matters, and tell the right person in language they can act on. Each job maps to a named component, and the authors derive a unified scoring function that ties the three together.
The first job sits in a Context Graph, a relational data structure that represents enterprise entities (contracts, deals, tickets, components) and the relationships and state transitions between them. In the reference implementation, that graph is built in Python with the NetworkX library, available in the paper's PDF. The graph is what makes each alert auditable, because each notification can point to the specific nodes and edges that changed rather than to a string that summarizes them.
Above the graph, the Delta Detection Engine monitors state transitions. Whenever a contract moves from review to signed, or a sales opportunity stalls past its expected close date, the engine produces a delta event that names the change and the graph path it touched. The event becomes a candidate for surfacing.
Not every delta deserves a notification. The Proactivity Scorer, the second architectural component, ranks each candidate through a single function the paper calls the Proactivity Score, which folds urgency, relevance, and persona-fit into one number. The authors derive the score formally in the paper body, treating the ranking problem as the formal core of what proactive surfacing means. A high-scoring event then reaches the Surfacing Layer, the third component, which uses Anthropic's Claude API to draft a notification grounded in the specific nodes that triggered it. The layer does not generate prose from scratch; each sentence ties back to the entity or transition that produced it, so the recipient can audit the chain.
The reference implementation ships as a complete Python codebase, which matters because anyone can run it on the case studies the authors chose rather than trust a vendor demo. Those case studies span three domains: contract lifecycle management, engineering incident response, and sales pipeline hygiene. The abstract reports Precision@5 of 0.83 and a false positive rate of 0.11 for the case studies, with mean time to surface reduced from 47 minutes to under 30 seconds — reported as the authors' own demonstration figures rather than an independently validated benchmark.
Two caveats sit on top of the result. The architecture ties the Surfacing Layer to Anthropic's Claude API, which means any team that runs the reference code inherits that vendor choice unless they swap the model client. And the paper itself is arXiv 2607.07721v1, a preprint without disclosed peer review or independent replication in the abstract. Treat the architecture as a formalism worth testing, not as a validated system.
The authors place current RAG and agentic frameworks on a single axis and call them fundamentally reactive, then position proactive surfacing as the next enterprise productivity step. Whether the rest of the field buys that argument depends on whether other labs can reproduce the architecture with a different model, score whether the Proactivity Score tracks what users actually want surfaced, and re-run the three case studies with disclosed metrics.
Next concrete read: whether an independent group ports the architecture off Claude and re-runs the same three case studies with a published metric.