diff --git a/backend/apps/agents/browser/browser_agent.py b/backend/apps/agents/browser/browser_agent.py
index a769dbd6..20df1926 100644
--- a/backend/apps/agents/browser/browser_agent.py
+++ b/backend/apps/agents/browser/browser_agent.py
@@ -1029,9 +1029,15 @@ async def run_browser_agent(
                 # result) for the third time in a row? If so, attach a loud
                 # warning to this tool_result so the model is forced to
                 # acknowledge it on its next turn.
-                call_key = _hash_tool_call(tu.name, tu.input, result)
-                is_loop = _detect_loop(recent_tool_calls, call_key)
-                if call_key[0] not in _LOOP_DETECTION_EXCLUDED_TOOLS:
+                # Loop detection only covers the non-excluded tools, so skip the
+                # hash entirely for the excluded ones; otherwise a screenshot/read
+                # serializes its full ~1MB result here just for _detect_loop to
+                # discard it (it short-circuits excluded tools to False anyway).
+                if tu.name in _LOOP_DETECTION_EXCLUDED_TOOLS:
+                    is_loop = False
+                else:
+                    call_key = _hash_tool_call(tu.name, tu.input, result)
+                    is_loop = _detect_loop(recent_tool_calls, call_key)
                     recent_tool_calls.append(call_key)
                     if len(recent_tool_calls) > _LOOP_WINDOW_SIZE * 2:
                         recent_tool_calls = recent_tool_calls[-_LOOP_WINDOW_SIZE * 2:]
diff --git a/backend/apps/agents/browser/browser_metrics.py b/backend/apps/agents/browser/browser_metrics.py
index 5c0ab342..836dc03d 100644
--- a/backend/apps/agents/browser/browser_metrics.py
+++ b/backend/apps/agents/browser/browser_metrics.py
@@ -54,7 +54,14 @@ def tier_for(tool_name: str) -> str:
     return _TIER.get(tool_name, "other")
 
 
+_metrics_dir_cache: str | None = None
+
+
 def _metrics_dir() -> str:
+    # Resolved + mkdir'd once, not on every tool call (this runs in the hot path).
+    global _metrics_dir_cache
+    if _metrics_dir_cache is not None:
+        return _metrics_dir_cache
     override = os.environ.get("OPENSWARM_BROWSER_METRICS_DIR")
     if override:
         base = override
@@ -69,6 +76,7 @@ def _metrics_dir() -> str:
         os.makedirs(base, exist_ok=True)
     except Exception:
         pass
+    _metrics_dir_cache = base
     return base
 
 
diff --git a/backend/tests/test_browser_hotpath_waste.py b/backend/tests/test_browser_hotpath_waste.py
new file mode 100644
index 00000000..3e2e28fd
--- /dev/null
+++ b/backend/tests/test_browser_hotpath_waste.py
@@ -0,0 +1,51 @@
+"""Hot-path waste removals in the browser sub-agent loop.
+
+Two per-action costs that were pure waste:
+  1. browser_metrics._metrics_dir() ran os.makedirs() on EVERY tool call.
+  2. The loop-detection hash serialized a tool's full result (a ~1MB screenshot
+     or 15KB read) even for tools that are excluded from loop detection, where
+     _detect_loop ignores the hash entirely. These pin both fixes.
+"""
+
+import os
+
+import backend.apps.agents.browser.browser_metrics as M
+from backend.apps.agents.browser.browser_loop import (
+    _detect_loop,
+    _LOOP_DETECTION_EXCLUDED_TOOLS,
+)
+
+
+def test_metrics_dir_is_cached_makedirs_runs_once(monkeypatch):
+    M._metrics_dir_cache = None
+    calls = {"n": 0}
+    real = os.makedirs
+
+    def counting(*a, **k):
+        calls["n"] += 1
+        return real(*a, **k)
+
+    monkeypatch.setattr(os, "makedirs", counting)
+    d1 = M._metrics_dir()
+    d2 = M._metrics_dir()
+    d3 = M._metrics_dir()
+    assert d1 == d2 == d3
+    assert calls["n"] == 1, f"makedirs must run once, ran {calls['n']}x"
+
+
+def test_excluded_tools_never_register_a_loop():
+    # The invariant the hash-skip relies on: for every excluded tool, even ten
+    # identical calls in a row are NOT a loop, so computing/storing the hash for
+    # them was dead work. Setting is_loop=False directly is therefore equivalent.
+    for tool in _LOOP_DETECTION_EXCLUDED_TOOLS:
+        key = (tool, "in", "out")
+        assert _detect_loop([key] * 10, key) is False, f"{tool} wrongly looped"
+
+
+def test_non_excluded_tool_still_loops_after_threshold():
+    # Guard the other side: the fix must NOT disable loop detection for the tools
+    # that need it (clicks/types/etc.).
+    key = ("BrowserClick", '{"selector":"#x"}', "clicked")
+    # below threshold -> not a loop; at/over threshold within the window -> loop
+    assert _detect_loop([key], key) is False
+    assert _detect_loop([key] * 5, key) is True