Token used (broken down to input token, output token, cached/not cached)
I have found that the agent hierarchy is already preserved locally, I can also find the number of requests/bubbles. But is there any extension that will give me a full observability on the agents + prompts.
We execute workflows. The orchestrator runs different sub-agents. We want to know which agent/sub-agent is consuming how much token, for which task.
Hey @Sheikh_Hasib_Ahmed! There’s no single view that shows agent → prompt → tokens per sub-agent yet, but you can assemble it:
Tokens per turn (incl. each sub-agent): every turn is its own usage event with a full token breakdown. Pull it from cursor.com/dashboard → Usage (CSV export), or via the Admin API POST /teams/filtered-usage-events (a team admin creates a read-only Team API key). Each row has tokenUsage (inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens) plus a conversationId. Admin API docs.
The join: that conversationId is the same chat/composer id you already have locally, and each sub-agent runs as its own conversation with its own id + parent link. Join the rows to your local hierarchy → orchestrator → sub-agent → tokens per turn.
Prompts + hierarchy: a subagentStart hook gives you task (the prompt), subagent_type, parent_conversation_id, and subagent_model. Hooks docs.
Cached vs not: cacheReadTokens is cached input, inputTokens is uncached, cacheWriteTokens is what was written to cache.
Having this joined into one view (with sub-agent labels) is an improvement we’re tracking, and I’ve added your use case to it.