Confirmed — the issue is solved for me too, on 3.10.x and later. I had the same problem on Windows 11 + WSL2 from mid-June through early July: agent tools (Shell, Grep, Glob) would fail with “Execution backend unavailable” while the integrated WSL terminal kept working fine.
What made it confusing was that it wasn’t one consistent failure mode. Sometimes only one chat was broken and a new chat fixed it. Sometimes the Editor agent panel failed but the Agents window worked for the same conversation. Sometimes every chat in a workspace was wedged until I renamed the project metadata folder under ~/.cursor/projects/. After an ungraceful reboot (laptop died without shutting down properly), several open chats failed at once; re-sending the last message from the Agents window usually recovered them. A full quit of Cursor (all windows) fixed the cases where multiple chats stayed broken overnight.
The integrated terminal working while agent tools failed was the key clue that this was Cursor’s agent execution path, not WSL or my shell setup.
I spent a fair bit of time debugging this with Cursor itself — ironically, the agent could help analyze its own failure mode when it wasn’t affected. We pulled evidence from agent transcripts, ~/.cursor-server logs (including RequestStore#acceptReply bursts after crashes), and Request Traces that showed shell dispatch hanging before reaching the extension host. I also cross-checked the forum threads, especially the remoteAuthority casing report (#164106), which matched what we were seeing.
Last confirmed failure for me was 3 July on 3.9.16. After upgrading past 3.10.11, I haven’t seen it again in about five weeks of heavy daily use across many workspaces. I did hit one ambiguous episode later (“no exit status” on trivial commands, no canonical error text, self-recovered within the session) — might be related, might not.
Workarounds that helped before the fix: Agents window for a wedged chat, new chat, full Cursor quit, and in the worst case resetting ~/.cursor/projects/<workspace-id>/. I would not pin an old version now; staying on 3.10.11+ is what fixed it for me.
Thanks to everyone who reported details here — especially the Jul 1 post on the same 3.9.16 build after a reboot, which matched my experience almost exactly.
And yes, this message was written with Composer 2.5, reviewed and approved by me-the-human 