Cursor shipped a context usage breakdown for agent sessions — you can now see exactly how your context window is being consumed during a run.
Context limits are one of the most common reasons AI coding agents silently degrade. When the window fills up, the model starts losing earlier instructions, file contents, and conversation history. Most tools give you no visibility into when this is happening. Cursor’s breakdown makes it visible before it becomes a problem.
TIP
If your agent starts producing repetitive or off-track code mid-session, check the context breakdown first. A nearly full context window is often the cause — starting a new session or trimming attached files usually fixes it.
This is most useful for developers running long multi-step agent tasks — refactors across many files, extended debugging sessions, or anything where the agent needs to hold a lot of state. Knowing which parts of the context are eating the most space (system prompt, file attachments, conversation history) lets you make smarter decisions about what to include.
The practical move: before starting a heavy agent session, open the breakdown after the first few steps to get a baseline. If you’re already at 60-70% context usage early on, trim your attached files or break the task into smaller chunks.