diff --git a/backend/apps/agents/browser/browser_agent.py b/backend/apps/agents/browser/browser_agent.py
index 64513aa8..45a2bf2c 100644
--- a/backend/apps/agents/browser/browser_agent.py
+++ b/backend/apps/agents/browser/browser_agent.py
@@ -935,7 +935,10 @@ async def run_browser_agent(
     # (measured: send done at turn ~11, then ~12 wasted perception turns). We drive
     # it to the OUTCOME and, if it keeps re-perceiving, end the run. A genuine
     # multi-send task issues its NEXT send (an action) which resets the stall, so
-    # only true spinning ends here.
+    # only true spinning ends here. This whole shortcut is meaningless for a
+    # gather/read task (no send to confirm), and arming it there let a cookie
+    # 'Accept all' click masquerade as the task's send, so we gate it on intent.
+    task_is_send = not deliverable_is_informational("", task)
     send_confirmed = False
     perception_stall = 0           # consecutive turns the model only LOOKED (no action)
     _POST_SEND_STALL_LIMIT = 2     # once the send registered, finish fast
@@ -1495,7 +1498,7 @@ async def run_browser_agent(
                 # across nodes, or scrolled off, so the text-probe is unreliable, which
                 # is exactly what left the model stalling to "double-check". A clean
                 # send click is proof enough; drive to the OUTCOME.
-                if not send_confirmed and "error" not in result and tu.name in _CONFIRM_TOOLS:
+                if task_is_send and not send_confirmed and "error" not in result and tu.name in _CONFIRM_TOOLS:
                     _cn = result.get("clickedName") or ""
                     _send_click = browser_batch_replay.is_replay_boundary(
                         {"action": "click", "name": _cn}) or any(
diff --git a/backend/tests/test_browser_agent_loop.py b/backend/tests/test_browser_agent_loop.py
index 505d5e3f..667a19e3 100644
--- a/backend/tests/test_browser_agent_loop.py
+++ b/backend/tests/test_browser_agent_loop.py
@@ -282,6 +282,27 @@ def test_run_that_never_calls_done_is_not_a_clean_success(monkeypatch):
     assert result.get("done") is False  # no explicit Done -> not a clean success
 
 
+def test_send_shortcut_does_not_arm_on_a_gather_task(monkeypatch):
+    # The Airbnb bug: a send-class click (here the index-99 sentinel = "Send", same
+    # as a cookie "Accept all" tripping the detector) on a FIND/gather task must NOT
+    # arm the send-completion shortcut, there is no send to confirm. If it did, the
+    # run cuts at the 2-turn post-send limit and leaks the canned "message went
+    # through" line. On a gather task it should run the full perception budget.
+    BH._browser_history.clear(); BH._domain_notes.clear()
+    primary = FakeLLM([
+        Resp([_rp("dismiss the cookie banner"), _tu("BrowserClickIndex", index=99)]),
+        *[Resp([_rp("keep reading the list"), _tu("BrowserScreenshot")]) for _ in range(8)],
+        Resp([_tu("Done", message="Here are the top items: a, b, c")]),
+    ])
+    aux = FakeAux()
+    _install(monkeypatch, primary, aux)
+    result = asyncio.run(BA.run_browser_agent(task="find me the top 10 repos", browser_id="b1", model="sonnet"))
+    # the send shortcut never armed: it ran past the 2-turn post-send cutoff toward
+    # the 6-turn perception budget, and no send-confirmation line leaked
+    assert primary.turn >= 6, f"gather task cut short at turn {primary.turn} (send shortcut wrongly armed)"
+    assert "went through" not in result["summary"]
+
+
 def test_early_perception_is_not_cut_short_before_any_action(monkeypatch):
     # Orienting on a cold/slow page can take several look-only turns; the stall
     # backstop must NOT fire before the agent has done anything (it only bounds a