diff --git a/backend/apps/agents/browser/browser_agent.py b/backend/apps/agents/browser/browser_agent.py
index 10fe71bd..8ee7b580 100644
--- a/backend/apps/agents/browser/browser_agent.py
+++ b/backend/apps/agents/browser/browser_agent.py
@@ -121,6 +121,7 @@ def _app_bridge_expression(tool_name: str, tool_input: dict) -> str:
 # reads poll briefly for the bridge to come up instead of declaring it absent.
 _BRIDGE_READY_WAIT_MS = 8000
 _BRIDGE_POLL_INTERVAL_MS = 400
+_bridge_known_absent: set[str] = set()
 
 
 def _parse_bridge_result(result: dict) -> object:
@@ -320,11 +321,22 @@ async def execute_browser_tool(
         # AppInvoke does not wait: its action either exists right now or it does
         # not, and a missing action should surface immediately.
         if tool_name in ("AppDescribe", "AppGetState"):
-            waited = 0
-            while waited < _BRIDGE_READY_WAIT_MS and not _bridge_ready(_parse_bridge_result(result)):
-                await asyncio.sleep(_BRIDGE_POLL_INTERVAL_MS / 1000)
-                waited += _BRIDGE_POLL_INTERVAL_MS
-                result = await _eval_once()
+            ready = _bridge_ready(_parse_bridge_result(result))
+            # Only the cold-boot read polls. Once a card is memoed bridge-absent,
+            # every later read takes the single eval above and skips the 8s sink;
+            # that single eval still detects a bridge that came up between reads.
+            if not ready and browser_id not in _bridge_known_absent:
+                waited = 0
+                while waited < _BRIDGE_READY_WAIT_MS and not ready:
+                    await asyncio.sleep(_BRIDGE_POLL_INTERVAL_MS / 1000)
+                    waited += _BRIDGE_POLL_INTERVAL_MS
+                    result = await _eval_once()
+                    ready = _bridge_ready(_parse_bridge_result(result))
+            # Record the verdict so the next read knows whether to pay the wait.
+            if ready:
+                _bridge_known_absent.discard(browser_id)
+            else:
+                _bridge_known_absent.add(browser_id)
         if tool_name == "AppDescribe":
             _persist_app_controls(browser_id, _parse_bridge_result(result))
         return result
diff --git a/backend/apps/agents/browser_agent_mcp_server.py b/backend/apps/agents/browser_agent_mcp_server.py
index 25a10606..b65543ae 100644
--- a/backend/apps/agents/browser_agent_mcp_server.py
+++ b/backend/apps/agents/browser_agent_mcp_server.py
@@ -118,12 +118,14 @@ TOOLS = [
         "name": "AppAgent",
         "description": (
             "Operate one of the user's OpenSwarm-built apps (a small web app they "
-            "created, e.g. a graphing or form app) that is open on the dashboard. A "
-            "dedicated app agent reads the app's own actions and state through its "
-            "native bridge and performs the task (no screenshots or DOM scraping), "
-            "then returns a summary plus a final screenshot. Use this for the apps "
-            "listed in the selected-app context, not for websites (use BrowserAgent "
-            "for those)."
+            "created, e.g. a graphing tool, a form, or a canvas game like Doom) that "
+            "is open on the dashboard. A dedicated app agent performs the task: it "
+            "drives the app through its native bridge when one is available (reading "
+            "the app's own state and calling its controls), and otherwise falls back "
+            "to native keyboard/mouse plus screenshots for canvas and game apps that "
+            "expose no bridge. It returns a summary plus a final screenshot. Works for "
+            "ANY app in the selected-app context, including games and canvas apps; use "
+            "BrowserAgent only for websites, not these apps."
         ),
         "inputSchema": {
             "type": "object",