Agent Terminal not working

Ditto.

Bug report: 30–60s agent response delay after commands (inference latency)

Product: Cursor IDE

Area: AI Agent / “Run command” workflow

Model: Sonnet 4.5, Composer 1, GPT-5.2, actually all of them

OS: Windows 10.0.26200

Shell: PowerShell (integrated terminal)

Workspace: on a windows dev drive (git repo)

Summary

After running terminal commands via Cursor’s AI Agent, the command finishes quickly but the agent UI remains in “Generating…” for ~30–60 seconds before responding. This delay appears to be LLM inference/response generation time (or queueing/time-to-first-token), not command execution or output processing.

Expected behavior

Agent should respond within a few seconds after command completion for trivial commands (e.g., echo ok), especially when terminal output/context is small.

Actual behavior

Command output appears immediately, but agent response arrives 30–60 seconds later.

Steps to reproduce

  1. Open Cursor on Windows.

  2. Open a project workspace (e.g., D:\MySweetCode).

  3. Start a new chat with Agent mode enabled.

  4. Open a fresh integrated PowerShell terminal (new terminal tab/window).

  5. Ask the agent to run:

echo ok

  1. Observe: terminal prints ok immediately, then the agent stays on “Generating…” for ~30–60 seconds before replying.

Repro frequency

High / consistent (observed repeatedly).

Diagnostics / what I tested

  • Created totally new PowerShell terminals in Cursor to eliminate scrollback/context bloat.

  • Repro still occurs even for trivial command output.

  • Symptom strongly suggests model call latency (queueing/time-to-first-token/thinking) rather than command runtime.

Impact

Severely slows development loop: even instant commands feel slow because the agent response lags by 30–60 seconds.

Attachments

  • Screenshot showing pwsh -NoProfile -Command “echo ok” outputting ok immediately while the agent remains “Generating…” (attached separately).

Request

Please investigate agent post-command latency with GPT-5.2 on Windows:

  • Is the agent waiting for additional background context gathering?

  • Is there excessive default “thinking/reasoning effort” applied post-command?

  • Is there queueing or throttling affecting time-to-first-token for GPT-5.2 in Cursor?

  • Any setting to reduce post-command agent latency for fast iteration loops?

yes. Cursor has been using the git bash terminal without issues until the upgrade to the new terminal system (or whatever prompted the need for a use Legacy Terminal setting), and then if I used the legacy terminal option it continued to work. Since the last update, maybe 2, even the legacy terminal keeps crashing.

I’m not sure how legit this is, but this issue with the terminal crashing seems to happen more often when commands output includes non-standard characters like emojis or colors. I’ve also noticed an increased occurrence when the agent tries to concatenate commands like

cd “c:myapp” && pint --dirty

This issue, along with the enormous expense of Cursor, pushed me to use Google Antigravity. Antigravity is massively cheaper and faster and better in every way really except for MCP support. Their mcp server support is buggy. Their terminal works way, way better. I trust it to do search and replace across multiple files simultaneously and edit multiple files in parallel. When it messes up, which is rare, it easily and quickly fixes it the majority of the time without me reverting. It is soooo much faster and more reliable. I also really like the memory system. It creates planning documents, tasks lists, and walkthroughs and indexes them for between thread recall. The mian thing is Cursor is just too expensive. It’s insane.

Not only is agent not working but it just logged me out of everything. Nice, LMAO

Unbelievable. Not only did Cursor agent stop working, it logged me out of everything, and now won’t open any folders. Completely unusable. Guys… what is going on?

Also, lol, I can’t even make a new topic in this forum.

Also all my chats are gone. LOL

The agent is going rogue, seriously destroying a bunch of my files in my codebase.

1 Like

I am simply sticking with Cursor version 2.1.42 until there’s confirmation here that this is resolved for Win11 / WSL2 users.

Considering switching to something else like https://kilo.ai/

1 Like

I didn’t see any mention of the terminal in the latest changelog but it seems to be a hell of lot more stable today. I’ve had zero terminal crashes this morning while working on two separate projects. So kudos to you guys for fixing this (if it actually was fixed and I’m not just getting lucky lol)

1 Like

I’m not entirely sure you’re having the same issue many of us are describing here, so I’m hesitant to try the latest version based on your experience.

Can anybody else confirm that the latest edition runs smoothly in WSL2 in Agent mode?

1 Like

@andrewh Can you confirm the above please?

I had a quiet day at work so installed the latest version:

```
Version: 2.3.29 (system setup)
VSCode Version: 1.105.1
Commit: 4ca9b38c6c97d4243bf0c61e51426667cb964bd0
Date: 2026-01-08T00:34:49.798Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Windows_NT x64 10.0.26200
```

Agent mode works fine in WSL2 and over SSH. Will report back if any major issues.

EDIT: Beware that the “Indexing & Docs” feature does not work, at least for the latest version.