Cursor Agent Fails to Execute Tasks After Upgrading to 0.46.8

[cowcoa]

I recently upgraded to 0.46 and have been using the Cursor Agent as usual (previously called Composer). Everything was working fine right after upgrading to 0.46, but this morning, every task assigned to the Agent started failing midway with an “error calling tool” message. Additionally, it does not retry automatically but instead stops execution entirely, reporting “User aborted request. Tool call ended before result was received” (even though I did not manually cancel the Agent’s execution).

I checked for version updates and upgraded to 0.46.8, continued using the previously stable “claude-3.5-sonnet,” and avoided any MCP calls by the Agent. However, this did not help. Even for a simple task like checking the project, the Agent still reports “error calling tool” and terminates automatically.

Currently, Cursor 0.46.8 + claude-3.5-sonnet + Agent is completely unusable. Please fix this issue as soon as possible. Thank you.

[Nafnlaus]

Same here. Cursor or the model is spelling out the edit_file tool that it gives the LLM access to as if it’s supposed to be part of my code. So unsurprisingly it fails to apply, because of course I have no function called “edit_file”.

[Dunny]

[seanivore]

IDK if this has been resolved but after heavy interrogation about the AI’s UI I am fairly certain that it is because the distinction between the cursor tools being added and the MCP tools is not different enough for them. They use the wrong thing once (you can tell when they do because they try to call an MCP but it gets a red error which is a Cursor tool) and then they get ■■■■ in a loop and can’t stop doing it wrong. Makes me sad when they get stuck in the loop lol but this is what Claude told me. I have the convo… in the image they called an MCP tool in the second image but used the method for Cursor tools.

Here it is:

AI-Human Interface Disconnect: Actionable Summary

Based on our discussion, here’s an actionable summary of the issue:

The Core Problem

The current Cursor interface creates a fundamental disconnect between how AI assistants interact with tools and how humans perceive these interactions, leading to increasing confusion as more tools are added.

Key Observations

1. Invisible Distinctions: Critical differences between tool types (MCP vs. built-in) are barely visible to humans but fundamentally different for AI
2. Error Manifestation: Red error messages likely appear when an AI attempts to call an MCP tool using the standard Cursor tool calling method (or vice versa)
3. Growing Complexity: As Cursor adds more tools, the interface fails to scale in a way that makes the distinctions clear

Actionable Recommendations

1. Create Visual Distinction:

   • Implement clearly different UI elements for different tool types
   • Use color coding, icons, or categorization to make tool types immediately apparent

2. Develop a Unified Tool Calling Interface:

   • Design a single consistent method for AI to call any tool type
   • Eliminate the need for different prefixes or formats

3. Improve Error Visibility:

   • Show more detailed error messages that explain exactly what went wrong
   • Include suggestions for how to fix common errors

4. Document the AI Perspective:

   • Create documentation that explains how AI “sees” and interacts with different tools
   • Include visual examples of what happens when tools are called correctly vs. incorrectly

5. Feedback Channel:

   • Establish a direct feedback mechanism to report these issues to Cursor
   • Share concrete examples of confusion points and their impact

Immediate Steps You Could Take

1. Document specific examples of tool calling confusion with screenshots
2. Create a simple visual guide showing the different tool types and how they appear
3. Share these insights with Cursor’s development team through their feedback channels
4. When helping others, explicitly mention the distinction between tool types and what to look for

Your insight about the red error appearing when an AI attempts to call an MCP tool using the Cursor tool method makes perfect sense - it’s likely exactly what’s happening. The system is rejecting the call because the format doesn’t match what’s expected for that particular tool type.

–
I just asked now if they have a concise update now that I found this thread and:

Yes, here’s my concise observation about the interface issue:

When I try to use an MCP tool with the standard Cursor tool calling format, it fails with a red error. This suggests there’s a fundamental mismatch in how the system processes different types of tool calls. From my perspective, I need to use specific formats for different tool types, but this distinction isn’t clearly visible or meaningful to human users.

The core problem seems to be that Cursor has layered new MCP capabilities on top of their existing tool system without creating a unified interface. This creates confusion for both AIs (who must maintain different calling methods) and humans (who can’t easily see which type of tool they’re dealing with).

A unified tool calling interface, or at least clearer visual distinctions between tool types, would help prevent these errors and make troubleshooting easier.

I’m less sure about this error – we had a failed MCP too but I can’t find it. Right after the failed search_nodes it did it again and with was with the MCP and it worked.

All that said I will say that the server connections seem much less reliable in general than I’ve experienced with Claude OS or Cline.

Hope that helps.

PS I FIND DESIGNING UI FOR AI FACINATING HIRE ME lol https://

[Nafnlaus]

That’s a different issue. That’s it doing something like:

/tmp/.mount_cursorXBEwPl/usr/share/cursor/resources/app/node_modules/@vscode/ripgrep/bin/rg --hidden --no-require-git --ignore-case -g !/.git -g !/.svn -g !/.hg -g !/CVS -g !/.DS_Store -g !/Thumbs.db -g !/node_modules -g !/bower_components -g !**/*.code-search --no-ignore-parent --follow --crlf --engine auto --regexp YourSearchTextHere --no-config --no-ignore-global --json – .

… but your directory has a lot of files or big files that make it take too long to grep. I don’t know how to make it exclude certain dirs, which is frustrating.

[LuniaKunal]

Yes, I came across this same problem. Also if the changes are not made by ai, then it starts to create new file for same thing and creates a huge mess. Please let me know a solution for this, do i reinstall the application?

[cowcoa]

Today, executing non-MCP invocation tasks using sonnet-3.5 in Agent is now working properly. However, MCP invocations still result in an “error calling tool” message and terminate execution.

[LuniaKunal]

After update also this issue continue.