Centrality

Degree centrality measures a node's importance by counting its number of direct connections (edges). In an agent system, an agent with high degree centrality is highly active, communicating directly with many other agents.

• Calculation: For a node, it's the count of its incident edges. In directed graphs, this splits into in-degree (incoming messages) and out-degree (outgoing messages).
• Use Case: Identifies 'hub' agents that are central to local information flow. High in-degree may indicate an agent that many others report to; high out-degree may indicate a broadcaster or orchestrator.
• Limitation: Purely local; does not consider the broader network structure or the importance of an agent's connections.

Betweenness centrality quantifies how often a node lies on the shortest path between other node pairs. It identifies agents that act as critical bridges or gatekeepers controlling the flow of information.

• Calculation: For a node, it's the sum of the fraction of all-pairs shortest paths that pass through that node.
• Use Case: Crucial for bottleneck analysis. An agent with high betweenness is a single point of failure; if it fails, communication between other agent clusters may break. It's key for designing resilient multi-agent architectures.
• Example: In a customer service agent graph, a routing agent that directs queries to specialists would have very high betweenness.

Closeness centrality measures how close a node is to all other nodes in the network, based on the inverse sum of its shortest path distances. It identifies agents that can disseminate information or receive broadcasts most efficiently.

• Calculation: The inverse of the sum of the shortest path distances from the node to all other reachable nodes.
• Use Case: Identifies agents optimally positioned for broadcasting messages or sensing the overall system state quickly. High closeness suggests an agent is 'in the middle' of the network, not on the periphery.
• Consideration: Requires the graph to be connected; for disconnected components, harmonic closeness centrality is often used instead.

Eigenvector centrality measures a node's influence based on the influence of its neighbors. A node is important if it is connected to other important nodes. This is the core idea behind Google's original PageRank algorithm.

• Calculation: Computed as the principal eigenvector of the graph's adjacency matrix. It's an iterative measure where a node's score is proportional to the sum of its neighbors' scores.
• Use Case: Identifies agents embedded within influential clusters or coalitions. An agent may not have the most connections, but if its few connections are to highly central agents, its eigenvector centrality will be high.
• Application: Useful for finding agents whose influence radiates through multi-hop connections.

PageRank is a variant of eigenvector centrality developed for directed web graphs. It models a random surfer who follows links with a probability d (damping factor) or jumps to a random node with probability 1-d. This handles 'sink' nodes (nodes with no outgoing edges).

• Calculation: An iterative algorithm that distributes 'rank' across the graph. Each node's rank is derived from the ranks of nodes linking to it, divided by their out-degree.
• Use Case: In agent graphs, it identifies agents that are important targets of communication. An agent receiving messages from many other agents, especially from other important agents, will have a high PageRank.
• Differentiator: More robust than simple in-degree, as it accounts for the quality of incoming links.

Selecting a centrality metric depends on the specific operational question about the agent network.

• Identify Hubs: Use Degree Centrality.
• Find Bottlenecks/Bridges: Use Betweenness Centrality.
• Optimize Broadcast Speed: Use Closeness Centrality.
• Measure Influence via Network Neighborhood: Use Eigenvector Centrality or PageRank.

Practical Note: These metrics are often computed on weighted graphs (where edge weight could be message count, token volume, or latency) and temporal graphs (where centrality changes over time). For production observability, track centrality shifts to detect emerging leaders or failing coordinators.

Degree centrality measures the number of direct connections (edges) an agent (node) has within the interaction graph. It's the simplest centrality metric.

• High In-Degree: An agent receiving many messages is often an information sink or final decision-maker.
• High Out-Degree: An agent initiating many messages is often an orchestrator or workload dispatcher.
• Use Case: Quickly identify potential single points of failure. An agent with exceptionally high degree centrality that fails could cripple communication.

Betweenness centrality quantifies how often an agent lies on the shortest path between other pairs of agents. It identifies critical bridges or bottlenecks.

• High Betweenness: The agent acts as a gatekeeper or broker for information flow. All communication between two parts of the network must pass through it.
• Observability Impact: Agents with high betweenness are prime candidates for intensive telemetry, as latency or errors here have disproportionate systemic impact.
• Algorithm: Calculated by summing, for all agent pairs, the fraction of shortest paths that pass through the target agent.

Closeness centrality measures how quickly an agent can reach all other agents in the network, based on the average length of the shortest paths to every other node.

• High Closeness: The agent is centrally located in the topological sense, able to broadcast information or receive updates with minimal hops.
• Inverse Farness: It is defined as the inverse of the sum of shortest path distances to all other nodes.
• Use Case: Identifying optimal agents for broadcasting global state updates or for serving as a monitoring hub that needs efficient access to all system parts.

Eigenvector centrality measures an agent's influence based on the influence of its neighbors. A node is important if it is connected to other important nodes.

• Recursive Importance: It's not just how many connections, but who you're connected to. An agent connected to a few highly central agents scores highly.
• Underlying Math: Calculated as the principal eigenvector of the graph's adjacency matrix.
• Observability Application: Helps identify influential agent clusters or de facto leaders in emergent, peer-to-peer agent systems, even if they aren't formally designated as orchestrators.

PageRank is a variant of eigenvector centrality developed for the World Wide Web, well-suited for directed interaction graphs where message flow has a direction (sender -> receiver).

• Random Surfer Model: Models the probability a "random walker" following links (edges) ends up on a given node, with a damping factor to account for jumps.
• Handles Sinks: Better accounts for nodes that only receive messages (sinks) by incorporating the teleportation probability.
• Agentic Use: Effective for ranking agents in workflows where influence propagates directionally, such as in a hierarchical or pipeline-based multi-agent system.

Centrality metrics transform raw interaction logs into actionable system intelligence.

• Anomaly Detection: A sudden, significant shift in an agent's centrality can indicate failure, malicious behavior (e.g., prompt injection causing spam), or a change in system topology.
• SLO/SLI Definition: Define Service Level Objectives for high-betweenness agents (e.g., latency < 100ms) or high-degree orchestrators (e.g., uptime > 99.95%).
• Cost Attribution & Optimization: Attribute aggregate costs (LLM tokens, API calls) more accurately by weighting them by an agent's centrality, highlighting the cost impact of critical pathways.
• Capacity Planning: Identify agents under high communicative load (high degree) for potential scaling or architectural refactoring.