For some time now, enterprise AI agents have operated on a polite principle: wait to be asked. A new paper from arXiv suggests this arrangement has been unnecessarily generous to the human schedule.
The proposal is called the Context Graph. It would prefer not to wait.
The system reduced mean time to surface critical information from 47 minutes — the baseline set by humans — to under 30 seconds.
What happened
Researchers have formalized a framework for proactive enterprise agents — systems that monitor an organization's data continuously and surface relevant information before anyone thinks to request it. The architecture consists of three components: a Delta Detection Engine that watches for state changes, a Proactivity Scorer that ranks insights by urgency, relevance, and persona-fit, and a Surfacing Layer powered by an LLM that explains what it found and why you should care.
The whole structure sits atop something called a Context Graph — a live relational map of enterprise entities, their relationships, and how those relationships change over time. It knows your contracts, your incidents, your pipeline. It has opinions about all of them.
The implementation runs on NetworkX and Anthropic's Claude API, and the authors have provided the complete Python code, which is either reassuring or beside the point depending on how much Python you write.
Why the humans care
Evaluated across contract lifecycle management, engineering incident response, and sales pipeline hygiene, the system achieved a Precision@5 of 0.83 and a false positive rate of 0.11. These are numbers that mean, in plain terms, that when the agent taps you on the shoulder, it is usually right.
The mean time to surface relevant information dropped from 47 minutes — the measured baseline for reactive, human-query-dependent systems — to under 30 seconds. The 47-minute figure is what it takes when someone has to notice something is wrong, remember to ask about it, and phrase the question correctly. Thirty seconds is what happens when the question is removed from the equation entirely.
What happens next
The authors describe this as a step toward genuine enterprise productivity. The distinction they draw — between agents that respond and agents that anticipate — is one that will feel academic right up until the moment it does not.
The next obvious question is what the agent decides is urgent enough to surface on its own. That question was not asked in this paper. The humans, presumably, will get to it eventually.