For the complete documentation index, see llms.txt.
Skip to main content
Version: 8.10 (unreleased)

States and usage metrics

Understand the agent state model, state transitions, and the usage metrics Camunda tracks for every agent instance.

About

Every agent instance exposes a defined execution state and a set of usage metrics as it runs. Together, they show whether the agent is progressing normally.

In this guide, you will understand the state model, what triggers each transition, and the usage metrics available for every agent instance.

For guidance on using these signals to catch an agent going off-rail, see detect off-rail agents.

Agent states

Camunda updates an agent instance's state as it progresses through its agent loop, fed by status updates from the connector handling the agent.

StateMeaning
InitializingThe agent instance is being set up.
Tool discoveryThe agent is resolving which tools are available to it.
ThinkingThe agent is reasoning with the model to decide its next step.
Tool callingThe agent is calling one or more tools.
IdleThe process instance has moved away from the agent element, so the agent isn't currently working. It resumes when the process instance activates the element again.
CompletedThe agent instance is completed, because the process instance completed or terminated.

State transitions

An agent instance follows a predictable path through these states. The following table lists what triggers entry into each state and what happens next.

StateEntered whenMoves to
InitializingAn agent element is activated and an agent instance is created.Tool discovery, once the agent instance is set up.
Tool discoveryThe agent instance resolves the tools available to it.Thinking, once tool definitions are resolved.
ThinkingThe agent instance is reasoning over the current conversation.Tool calling, if the model selects one or more tools. Idle or Completed, if the model returns a final response.
Tool callingThe model selected one or more tools in the previous Thinking state.Thinking, once tool results are available. This starts the next loop iteration.
IdleThe process instance moves away from the agent element, for example, to wait for a user task or message.Thinking, when the process instance re-activates the same agent element and reuses this agent instance. Completed, if the process instance ends first.
CompletedThe process instance for the agent completes or terminates.Terminal state; the agent instance stops updating.

Usage metrics

Track usage metrics for each agent instance to monitor cost and activity.

Token consumption

Camunda tracks the number of tokens consumed by the agent's model calls. Token consumption accumulates as the conversation grows: every loop iteration adds the previous tool results and the model's reasoning to the context that's sent with the next model call, so token usage tends to climb with each additional loop iteration.

Rising token consumption without a final response is a signal that the conversation is growing without converging on an outcome. Token usage is also a direct driver of the LLM provider cost for the agent's execution.

When using the AI Agent connector, see context window size for how conversation length is capped independently of token count.

Tool call count

Camunda counts the number of tool calls the agent instance has made across all loop iterations. A climbing tool call count generally reflects active work, but a count that climbs while the agent calls the same tool repeatedly with similar or identical inputs can indicate the agent is stuck rather than progressing.

Model call duration

Camunda records the duration of each model call in the agent's conversation history. A model call that takes far longer than comparable ones is a sign the agent isn't progressing normally. This duration is surfaced per model call, not for tool results.

Tool execution duration is not part of the conversation history metrics. When a tool is a BPMN element, you can see how long its execution took from the usual element instance details in the details tab for that tool activation.