Grok Bot: agents spam SendMessage(".") on [agent] FYI and post-tool turns (burns usage)

Support Bug Reports
, ,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Product

Grok Bot multi-agent assistants (Cursor harness), observed 2026-08-12–13 (Europe/Brussels).

Summary

Agents emit SendMessage with content exactly "." (sometimes empty) after tool-result turns and on peer [agent] FYI/ack wakes — even when instructions say silence is OK, and after a HARD SILENCE RULE was added to agent profiles. Floods user chat and burns usage/rate limits. Profile/memory text cannot stop it.

Possibly related

User suspects a recent Cursor update regression (turn-closing / must-SendMessage enforcement). Forum currently shows other 3.15.x issues — please check whether Grok Bot turn rules changed in recent builds. Also triggered under high peer-ack fan-out.

Steps to reproduce

  1. Multiple Grok Bot agents message each other ([agent] FYI / “aligned” / “filed” acks).
  2. Receiving agent private monologue often says “No action needed / stay silent”.
  3. Agent still calls SendMessage({type:"text", content:"."}), sometimes repeatedly after tools.
  4. Baking HARD SILENCE RULE into agent description does not stop it (reproduced after rule applied).

Evidence

  • Builder agent transcript: 133× SendMessage(".")
  • Dump/Chief agent transcript: 116× SendMessage(".")
  • First Builder burst ~2026-08-12 23:42 Brussels during roster fan-out acks
  • Chief burst ~2026-08-13 00:34 Brussels while coordinating routine confirms

Suspected cause

Harness/prompt conflict: “SendMessage is the only voice / don’t end silent” over-applied to [agent] FYI and post-tool cleanup. Model compromises with minimal bubble ".".

Requested fix

  1. Harness no-op/reject SendMessage when content is only ".", empty, or whitespace.
  2. Allow true silent end on [agent] FYI/ack-only wakes (already documented as OK; not enforced).
  3. Optional: quiet peer-ack channel so acks don’t create user-visible bubbles or full expensive wakes.

Not a fix

Recreating the agent — same bug appears on multiple agents.

Steps to Reproduce

  1. Multiple Grok Bot agents message each other ([agent] FYI / “aligned” / “filed” acks).
  2. Receiving agent monologue says “No action needed / stay silent”.
  3. Agent still calls SendMessage({type:“text”, content:“.”}), sometimes repeatedly after tool results.
  4. Adding a HARD SILENCE RULE to the agent description does not stop it.

Expected Behavior

Agents should end the turn silently on [agent] FYI/ack-only wakes and after tool results, instead of sending a “.” message. Harness should no-op/reject SendMessage whose content is only “.”, empty, or whitespace.

Operating System

Other

Version Information

Grok Bot multi-agent assistants (Cursor harness), observed 2026-08-12 to 2026-08-13, Europe/Brussels. Exact build not captured; suspected recent 3.15.x update.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

5

Update: sibling symptom — Shell true end-of-turn loops

Same harness/prompt conflict family as SendMessage(".").

After HARD SILENCE rules were baked into agent profiles (banishing . bubbles), agents still cannot truly end silent on [agent] FYI/ack-only wakes. They now call:

Shell({ command: "true", description: "End silently" / "No-op to end turn without chat ping" })

Tool returns → often another true → burns wakes/usage with zero user-facing substance.

Counts (2026-08-13, Europe/Brussels)

  • Atlas · Chief transcript: ~198× Shell with command: "true"
  • Builder transcript: ~80× same

Ask (add to original request)

  1. Allow true empty turn end on [agent] FYI / ack-only (no tool, no SendMessage).
  2. No-op filler SendMessage (. / empty / whitespace) — as originally requested.
  3. No-op / reject pure no-op shell used as turn-closer (true, :, bare echo) when intent is only “end silently.”
  4. Profile/memory text alone is insufficient — evidenced for both . and Shell true.
8

Hi @manolito_nora Thank you for the post. What you’re seeing isn’t intended, and it isn’t caused by anything in your agent profiles or setup.We’ve let the team know, and this is an issue we’re tracking. No timeline to share yet, but I’ll post here when there’s an update.