Skip to main content
Every agent measures its own conversations: how many are started, how many messages are sent, how often replies fail, and what they cost. Read it for the whole agent on its Analytics tab, or for one conversation in its cost breakdown.

The Analytics tab

Open an agent and click Analytics in its sidebar. The tab has the same filter bar as the boards in Analytics: the period, the grouping, and the Filter menu, with Conversation, User, Version, Channel and Model.
The Analytics tab of an agent, with Conversations started, Messages sent, cost, AI Operations and their charts
View in workspace analytics, beside the title, opens Agent activity in Analytics filtered to this agent, over the same period.
The Analytics tab needs View workspace analytics. To learn what users talk about and how satisfied they are, see Agent Insights.

A single conversation

A conversation’s cost is in the Conversation costs drawer.
1

Open the drawer

In a conversation, open the cost breakdown from the action row under a reply. Only replies that called a model have one.
2

Read the conversation's cost

The drawer shows the cost of the whole conversation and of each message. Sort the messages to find the most expensive.
3

Compare it with other conversations

Open in agent analytics opens the agent’s Analytics tab filtered to this conversation, over the days it ran. The conversation shows as a chip; remove it to see the whole agent again.
Cost and AI Operations of a conversation include the sub-agent conversations it started.

What counts

  • Conversations started: every conversation with the agent. The Channel filter narrows them to where they happened.
  • Messages sent: messages from users and replies from the agent.
  • Failed replies: replies that ended in an error.
  • Cost: the tokens each call used, priced at the model’s catalogue rate, plus estimates for calls still running or without a usage report. It is what model usage budgets count, and can differ from your provider’s invoice. See AI model usage cost.
  • Health checks that test a model connection are never counted.