Source-linked AI summary
Agentic Network Traffic Monitoring
Manuel Tsoukatos, Hayden Jananthan, Jeremy Kepner
TL;DR
Agentic systems bring security risks, making it important to monitor agent behavior through its network traffic. This paper combines DBOS, OneSparse, and GraphBLAS in a simulator that records communications as complex-valued traffic matrices. The resulting matrices distinguish communication patterns across exploration scenarios, including routed messages and dropped packets.
Problem
As agent use expands, security vulnerabilities such as prompt injections, data leaks, and adversarial actors make auditing and monitoring agent behavior important.
Source: §I · p. 1Method
The paper combines DBOS, OneSparse, and GraphBLAS in a multi-agent exploration simulator and records successful and unsuccessful communications in complex-valued traffic matrices.
Source: §I · p. 1, §II · pp. 1–2Results
The simulated scenarios produce distinct traffic patterns: daisy chaining routes messages through intermediaries, while a five-square range setting produces many dropped packets.
Source: §III · p. 4Takeaways & Limitations
The traffic matrices provide a way to visualize agent connections and serve as a foundation for monitoring agent behavior and detecting anomalous or adversarial behavior.
Source: §III · pp. 2–3
Abstract
from arXiv · showhide
As the use of agentic artificial intelligence increases in nearly every industry, there exists a widening attack surface. It is necessary to monitor agents to ensure that agents are acting in a way that is aligned with the users intent. Auditing an agent's network traffic provides a clear record of the agent interactions. This work presents a novel approach to monitoring the network traffic of agentic systems using complex valued hypersparse traffic matrices by integrating DBOS (DataBase OS), the OneSparse PostgreSQL database, and the GraphBLAS math library. To develop these concepts an agentic simulator was constructed, allowing a varying numbers of AI agents to collectively survey a virtual environment using different strategies. The resulting network traffic matrices enable easy monitoring of the AI agents.
I. INTRODUCTION
Agentic systems create security risks that make auditing and monitoring behavior important, and network traffic can provide insight into agent behavior. The paper combines database-backed workflows with hypersparse traffic matrices to monitor agents without internal knowledge of them.
I. INTRODUCTION
- Network traffic analysis can provide insight into agent behavior because traffic records data flowing between sources and destinations.The passage cites prior work on this connection,.Source: §I · p. 1
- Traffic matrices encode packet counts from source rows to destination columns, and hypersparse storage suits large matrices whose entries are mostly zero.GraphBLAS supports creating and operating on hypersparse matrices; OneSparse binds GraphBLAS matrices to PostgreSQL as an SQL table type.Source: §I · p. 1
- DBOS stores workflow events as structured database data and provides durable execution guarantees, including completion despite crashes and logically correct reverse actions.The paper presents these features as useful for tracking long-running, complex agentic workflows.Source: §I · p. 1
- Combining DBOS, GraphBLAS, and OneSparse, the approach uses traffic matrices to monitor agent behavior without internal knowledge of the agents.The work applies these technologies in a simulator where varying numbers of agents survey a virtual environment using different strategies.Source: §I · p. 1
II. METHOD
The method simulates multi-agent exploration on a configurable grid and records agent communications through a database-backed network interface. It aggregates successful and unsuccessful message traffic into two complex-valued matrices, with scenarios varying communication range, routing, delegation, and agent control.
II. METHOD
- The simulator places an orchestrator on a configurable two-dimensional grid, where subagents explore to map the ground-truth environment.The grid combines impassable walls and passable space, and the orchestrator relies on subagents to survey it.Source: §II · p. 1
- Subagents observe their immediate three-by-three surroundings, maintain local maps, and report findings to the orchestrator, which assigns new destinations.Agents interact with the ground-truth database through DBOS steps when moving and observing.Source: §II · pp. 1–2
- A simulated network interface logs communication in PostgreSQL, recording delivery through DBOS.send and outcomes through DBOS.step for later analysis and live monitoring.The orchestrator communicates with subagents only through the custom networking function.Source: §II · p. 2
- Two complex-valued GraphBLAS matrices distinguish end-to-end message counts from intermediary hops, with real values for delivered messages and imaginary values for unsuccessful attempts.The matrices are stored using OneSparse; the hop matrix differs when messages pass through intermediaries.Source: §II · p. 2, Figure 2 · p. 2
- The baseline uses a heuristic that assigns the nearest unknown position to the nearest ready subagent, assuming unlimited-range, two-way communication.Subagents use breadth-first search over locally known places when given an order.Source: §II · p. 2
- The testbed varies peer updates, range limits, daisy-chain routing, hierarchical delegation, and LLM-based reasoning instead of simple heuristics.These options allow different communication and agent-control scenarios to be simulated.Source: §II · p. 2
III. RESULTS
The simulations show distinct communication patterns across baseline, daisy-chain, hierarchical, and low-range daisy-chain scenarios. The resulting traffic matrices are presented as a basis for monitoring agent behavior and detecting anomalous or adversarial behavior.
- The traffic matrices provide a way to visualize agent connections and are presented as a foundation for monitoring behavior and detecting anomalous or adversarial activity.Source: §III · pp. 2–3
- In the baseline scenario, true source-destination and all-hop matrices look identical, while the expected pattern has nonzero values only in the first row and column.Source: §III · pp. 3–4, Figure 3 · p. 3
- With a 20-square Manhattan range limit, daisy-chain traffic retains the baseline end-to-end matrix shape, but the intermediary-hop matrix diverges as messages route through closer subagents.The scenario completed with the mental map at 100%.Source: Figure 3 · p. 3, §III · p. 4
- In the hierarchical scenario, only two subagents interface with the orchestrator, and the matrices show the subordinate communication chain.Source: Figure 3 · p. 3, §III · p. 4
- A 5-square range limit in the low-range daisy-chain scenario produces many dropped packets as subagents struggle to relay messages.The passage identifies dropped-packet counts as a quantity that could be minimized when exploring connectivity approaches.Source: Figure 3 · p. 3, §III · p. 4
IV. CONCLUSION AND FUTURE WORK
The paper combines DBOS, PostgreSQL, and GraphBLAS to create a customizable simulation for exploring agentic communication patterns. Future work includes expanding scenarios, generating datasets, and adding network monitoring to current agentic systems.
IV. CONCLUSION AND FUTURE WORK
- DBOS and PostgreSQL preserve state and manage workflows, while GraphBLAS models traffic in a customizable simulation for exploring agentic communication patterns.Source: §IV · p. 4
- Experiments across thousands of runs and subagents could provide useful data on agentic behavior and support datasets of normal and anomalous network patterns.Source: §IV · p. 4
- Future work may add network monitoring to current agentic systems and represent delivered-versus-dropped packet phase in a real-valued matrix.Source: §IV · p. 4