diff --git a/backend/apps/agents/browser/browser_agent.py b/backend/apps/agents/browser/browser_agent.py
index 5297b105..2d72e15d 100644
--- a/backend/apps/agents/browser/browser_agent.py
+++ b/backend/apps/agents/browser/browser_agent.py
@@ -1178,14 +1178,18 @@ async def run_browser_agent(
             browser_history.prune_stale_page_state(messages)
             browser_history.place_cache_marker(messages)
             p_llm_t0 = time.monotonic()
-            response = await p_cancellable(client.messages.create(
-                model=api_model,
-                max_tokens=4096,
-                # Cache the ~4k-token fixed prefix (system + tool schema) so it's reprocessed once, not on every turn: big TTFT + cost win on the first run, which is dominated by turns x per-turn prefill. The trailing cache_control marker is what Anthropic keys on; on non-Anthropic routes (9router) the marker is harmlessly ignored.
-                system=p_cached_system,
-                tools=p_cached_tools,
-                messages=messages,
-            ))
+
+            # STREAM, don't .create(): 9Router returns non-Anthropic lanes (Gemini/OpenRouter/Antigravity) as a REAL multi-event SSE stream that the non-streaming client parses to empty content, silently breaking tool-use on every non-Claude provider (measured: only cc/ emitted tool_use before this). The streaming parser reconstructs tool_use identically for ALL providers, so this is the model-independence fix, not a UX tweak. Cache marker still rides p_cached_system (Anthropic keys on it; other routes ignore it harmlessly).
+            async def p_stream_turn():
+                async with client.messages.stream(
+                    model=api_model,
+                    max_tokens=4096,
+                    system=p_cached_system,
+                    tools=p_cached_tools,
+                    messages=messages,
+                ) as p_s:
+                    return await p_s.get_final_message()
+            response = await p_cancellable(p_stream_turn())
             if response is None:
                 break
             p_llm_ms = int((time.monotonic() - p_llm_t0) * 1000)
diff --git a/backend/tests/test_browser_agent_loop.py b/backend/tests/test_browser_agent_loop.py
index df46d21c..eb6fcefe 100644
--- a/backend/tests/test_browser_agent_loop.py
+++ b/backend/tests/test_browser_agent_loop.py
@@ -30,17 +30,32 @@ class Resp:
         self.usage = type("U", (), {"input_tokens": 1, "output_tokens": 1})()
 
 
+class FakeStream:
+    # mirrors anthropic's messages.stream(): async CM whose get_final_message() returns the turn
+    def __init__(self, resp): self.resp = resp
+    async def __aenter__(self): return self
+    async def __aexit__(self, *a): return False
+    async def get_final_message(self): return self.resp
+
+
 class FakeLLM:
     def __init__(self, scripted):
         self.scripted = scripted; self.turn = 0; self.calls = []
         self.messages = self
 
-    async def create(self, **kw):
+    def p_next(self, kw):
         self.calls.append(kw)
         i = min(self.turn, len(self.scripted) - 1)
         self.turn += 1
         return self.scripted[i]
 
+    async def create(self, **kw):
+        return self.p_next(kw)
+
+    def stream(self, **kw):
+        # the loop now streams; return an async-CM yielding the scripted turn
+        return FakeStream(self.p_next(kw))
+
 
 class FakeAux:
     def __init__(self):