Does Cursor’s Agent consume tokens while in an idle/waiting state?
Specifically:
If the agent is not actively sending prompts or generating responses, but is just waiting (idle), are tokens still being used during that time?
Or are tokens only consumed during actual model activity such as prompting, generation, or reasoning?
Also, are there cases where the agent appears idle but is still consuming tokens in the background?
No, the agent does not consume tokens while waiting for a shell command to finish. What you’re seeing (“Waiting 14m 24s for shell”) means the agent issued a terminal command and is waiting for it to complete. No model calls are happening during that wait, so no tokens are used.
Tokens are only consumed when the AI model is actively working: reading your prompt, reasoning, generating text, or deciding which tool to call next. The cycle looks like this:
Model thinks and responds (tokens consumed)
Tool executes (e.g., shell command, file read, search) — no tokens consumed
Results sent back to the model (tokens consumed when the model reads and processes the results)
So in your screenshot, the agent is in step 2 — waiting for a tool to finish. Your usage is not increasing during that time.
There is one background case worth knowing about: if a conversation gets very long, the agent may run a one-time conversation summarization to compress the history. This does use a small amount of tokens, but you’d see a “Summarizing conversation” indicator, and it only happens when the context window is nearly full. It’s not related to idle waiting.
You can always check your real-time usage and token breakdown on your dashboard.