Regression? Agent Context Compacted Loop with Large Responses

Support Help
1

Previously, when we get a large response from MCP server, cursor was able to write the large output to ~/.cursor/projects/… and will be smart enough to write a python script to analyze the content to complete the task.

However, it looks like with recent updates - this is no longer happening and instead the agent will try to summarize it but ends up with “Agent context compacted”. It will than try again and again.

I am not sure if this is due to the recent change in https://cursor.com/blog/dynamic-context-discovery

2

Hey, thanks for the report.

This looks like a known issue with handling large outputs. The team is working on a fix using dynamic context discovery https://cursor.com/blog/dynamic-context-discovery.

As a temporary workaround, you can ask the agent to first save the large MCP output to a file, then analyze it. For example:

"save the output to temporary/output.txt, then analyze the file contents"

To help us understand this better, could you share:

  • Your Cursor version (Help > About)
  • The Request ID from the chat where this happens (chat context menu > Copy Request ID)
  • Which MCP servers you’re using
  • The approximate size of the MCP response that triggers the loop
3

Thanks for the quick response i don think asking it to write the response to the output helps as it looks like the agent is attempting to read the context even before it checks the size of the responses.

this is an internal remote mcp.

Below are two screenshots-

i took one a few months back (nov - cant remember the version, im now on beta with nightly buidls)

the 2nd was taken recently. i hope you will find this useful

redacted
redacted1870×725 32.1 KB

(top able to write to file)

failedreduction
failedreduction961×759 23.7 KB

1 Like
4

i think this might be related

I thought this was fixed and was super excited. we have a wide range of tools here and so far cursor seems to be the only one that “solved” this - at least it was a few months back. but now it is back to the same issue.

1 Like
5

Interestingly no issue with claude code as it will store the tool results in a temporary file - i think this was what i was observing with cursor previously