# Azure AI Agent

> Azure AI agents in the Foundry Agent Service carry no meter of their own. Every conversation bills through the model deployment and tools the agent invokes. ZopNight discovers each agent with its model-deployment linkage and groups them under the owning account, making forgotten agents and the deployments that still bill for them visible.

Source: https://zop.dev/integrations/azure/ai-agent
Updated: 2026-08-19

---

Azure AI agents are configured assistant instances in the AI Foundry Agent Service, combining a model deployment with tools and instructions. Their cost flows through the model and tool calls they generate.

## An agent's bill arrives indirectly

Creating an agent is free. Running one is not. Each conversation turn consumes tokens on the model deployment the agent is bound to, and each tool invocation (search, code execution, external APIs) bills wherever that tool meters. So the agent object is best read as a cost multiplier attached to a deployment: an agent wired into a busy application can dominate a deployment's token volume while never appearing on the invoice under its own name.

## Linking agents to the deployments that bill

Discovered via the AI enricher with model-deployment linkage, mapping agent workloads onto the deployments that bill for them. ZopNight groups the agents of each Cognitive Services account under a synthetic Agents node, keeping the account as grandparent so the chain from agent to deployment to billed account stays navigable. Agents are topology-only rows (nothing to stop or schedule), but they answer the question a raw deployment list cannot: what is actually driving traffic to this model.

## Agent estates that drift

Agent inventories accumulate in a way deployments do not, because agents are cheap to create and invisible when idle. Common findings: prototype agents from a demo still bound to a production deployment; agents whose instructions reference retired tools or data sources, failing silently in ways that still consume tokens on retries; and multiple near-identical agents created by different team members against the same deployment, making usage attribution guesswork. A periodic sweep of agents against their linked deployments keeps all three in check.

## Finding agents in the Agent Service

The AI Foundry portal's Agents section lists every agent in the selected project, with the model deployment, tools, and instructions each one carries. Agents do not surface as first-class rows in the Azure portal, which is precisely why an inventory that expands them from the account side is worth having.