Frequent "Agent Execution Timed Out" errors during agent chats

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

CURSOR SUPPORT REPORT — “Agent Execution Timed Out” (deadline_exceeded)

SUMMARY

Frequent “Agent Execution Timed Out” errors during agent chats. The request
connects successfully, but the server-side agent backend sends ZERO tokens and
the request fails at a consistent ~70-second deadline (“Request aborted before
first token”). This happens across chats of all sizes (13-message chats fail as
often as 187-message chats) and affects roughly 35% of requests
(33 success / 18 timed out / 6 cancelled across recent sessions).

All client-side network causes have been investigated and ruled out (details
below), which points to a server-side agent-backend issue. Requesting
investigation of the request IDs listed.

ENVIRONMENT

  • OS: Windows 11 (10.0.26200), x64
  • Cursor client_version: 3.12.30 (stable channel)
  • Layout: glass
  • Model observed on failures: claude-opus-4-8
  • Backend: server-side “agent backend” (useAgentProviderRouter=false)
  • Network: home Wi-Fi, no proxy (WinHTTP direct, WinINET ProxyEnable=0)

ERROR SIGNATURE (representative trace, requestId f79fc100-…)

19:32:09 agent.turn.start turn_type=“resume” model=“claude-opus-4-8”
19:32:09 Starting stream request → Using agent backend
19:32:10 “No first token received within 2s”
19:32:12 …4s → …6s → …8s … → …32s (server sends nothing)
19:33:18 [error] ConnectError: [deadline_exceeded] Agent Execution Timed Out
19:33:18 “Request aborted before first token” elapsedMs=70136
span “agent.request” durationMs=70299

All 18 failures share this pattern; elapsed time clusters tightly at
~70,000-71,000 ms (a couple at 77-81s), i.e. a fixed server-side deadline with
zero bytes received. TCP connection stays up the whole time (no resets).

CLIENT-SIDE INVESTIGATION (all clean)

  • Extension host: starts cleanly every session; cursor-agent-exec /
    cursor-agent-worker / cursor-mcp all activate in <1s. No startup timeout.
  • DNS: api2.cursor.sh + all Cursor hosts resolve fast (~17ms). (Note: an earlier
    family-filter DNS resolver 1.1.1.3 was causing intermittent ENOTFOUND; switched
    to 1.1.1.1 — resolved.)
  • HTTP transport: set cursor.general.disableHttp2=true (HTTP/1.1). Confirmed live:
    “HTTP/1.1 transport created”, httpVersion=1.1. (Earlier HTTP/2 produced 122x
    NGHTTP2_REFUSED_STREAM; gone after switching to HTTP/1.1.)
  • Endpoint reachability (curl): api2.cursor.sh dns=17ms tls=195ms ttfb=259ms HTTP 200.
  • Stream buffering test (curl to httpbin drip): first byte in ~0.3s, streams
    incrementally over 6s — no middlebox buffering on this host.
  • Endpoint security: SentinelOne agent (Sentinel Agent 25.1.3.334) with
    SentinelNetworkMonitor loaded, but logged ZERO network/mitigation events at any
    timeout timestamp, and generic HTTPS streaming passes through fine.
  • Windows event logs: NO TCP drops, Wi-Fi disconnects, or reconnects during the
    timeout windows.

Conclusion: DNS + HTTP/1.1 fixes eliminated the earlier connect/stream errors,
but “Agent Execution Timed Out” persists after those fixes and after a full
restart. Given the clean client evidence and the zero-token / fixed-70s-deadline
signature on the server-side agent backend, this appears to be a server-side
agent execution timeout.

FAILING REQUESTS (18) — timestamp (local, UTC-7) | requestId | conversationLength

2026-07-21 22:22:09 97055414-c5be-46b1-b91e-317b387d3a30 len=18
2026-07-21 22:23:46 09f46f58-1783-4949-b1f8-b4717708bd4f len=18
2026-07-22 13:04:20 091628c1-7070-4508-b40b-754968a42b92 len=20
2026-07-22 13:08:55 2f014a94-a413-425a-898e-7c68ed2f3779 len=20
2026-07-22 14:56:02 ee8162a2-a06a-4827-aaab-adce6e37a3b2 len=13
2026-07-22 14:58:04 0b588674-5368-446d-8a48-32e3ec16b5a4 len=13
2026-07-22 15:11:38 03d0a3a8-4714-442d-bbaa-814f711adb12 len=13
2026-07-22 15:14:14 b01ba0a6-062b-4411-bab7-34e548c07c04 len=13
2026-07-22 17:44:17 7abe0216-3e72-4aa4-b298-e22eeafc06b0 len=13
2026-07-22 17:46:24 1ec64755-eb65-483f-839d-66bc12d999eb len=13
2026-07-22 18:51:14 cd05f662-acf8-4822-8f7f-a87a0426bf2b len=187
2026-07-22 18:52:46 6b3c8418-e01b-429e-96ca-20bd373b5ca7 len=187
2026-07-22 18:54:06 a141b79a-a1a4-40b5-98c0-d704f5d52fe9 len=187
2026-07-22 18:56:23 128a7be1-c58a-4d0c-b5c1-3500f01e09f5 len=13
2026-07-22 19:01:57 ab477ad9-a861-49a3-92b4-a5f3ce401b3e len=13
2026-07-22 19:03:47 eb78ab48-afb8-4804-bb9d-49f55b6eb34f len=13
2026-07-22 19:05:57 dd49df84-a3b2-491b-9c83-28db458b4d4a len=187
2026-07-22 19:33:18 f79fc100-30be-4a7c-a1ec-ead488610512 len=187

Most recent Request ID (from the Cursor error UI): f79fc100-30be-4a7c-a1ec-ead488610512

QUESTION FOR SUPPORT

Why is the server-side agent backend failing to emit a first token within the
~70s deadline for these requests, and is there anything server-side (routing,
model capacity for claude-opus-4-8, account/region) that explains the ~35%
failure rate? Client network path is confirmed healthy.

Steps to Reproduce

Occurs intermittently during agent chats

Operating System

Windows 10/11

Version Information

Cursor client_version: 3.12.30 (stable channel)

For AI issues: which model did you use?

claude-opus-4-8

For AI issues: add Request ID with privacy disabled

see IDs above

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

Hey there, thanks for the detailed write-up.

Quick but important correction: this isn’t a server-side issue. I checked all 18 request IDs and none of them reached our servers - the timeout happens locally while Cursor waits for the agent execution provider (the extension host) to start, so the request never gets sent. That’s also why your network fixes didn’t help: the failure is before the network.

What should unblock you:

  1. When it hits: Ctrl+Shift+P → Developer: Restart Extension Host (faster than a full reload, and usually clears it).
  2. For a stuck conversation: start a fresh chat and carry context over with @Past Chats — your own data shows new chats work while the same threads keep failing.
  3. Durable fix (Windows + SentinelOne): have IT add endpoint-security process and path exclusions for Cursor (%LOCALAPPDATA%\Programs\cursor\ and %APPDATA%\Cursor\). Security software scanning file I/O at startup is the usual cause on Windows. Guide: Endpoint Security Configuration.

If it keeps happening after that, grab the Extension Host log (View → Output → Extension Host) from around a failure and let me know whether Restart Extension Host clears it — that’ll pin down your variant.