[eric] browser: steer the model to page URLs, not raw JSON/API endpoints (read APIs via BrowserReplayRoute)

backend/apps/agents/browser/browser_fast_path.py

@@ -65,7 +65,9 @@ P_CLASSIFIER_SYSTEM = (
     "If line 1 is READ or ACT, follow it with a short browsing brief:\n"
     "ENTRY: the best starting URL; use a direct deep/search URL when the site's "
     "pattern is well known (LinkedIn people search is "
-    "https://www.linkedin.com/search/results/people/?keywords=NAME).\n"
+    "https://www.linkedin.com/search/results/people/?keywords=NAME). Always a normal "
+    "page a person would see, never a raw JSON/API endpoint (e.g. Instagram's "
+    "web/search/topsearch), which renders as an unreadable data wall.\n"
     "Then 3-6 numbered steps, one short action each.\n"
     "Copy any text the user wants typed, sent, or posted EXACTLY, character for "
     "character. Never invent names, values, or wording the user did not give."

backend/apps/agents/browser/browser_schema.py

@@ -243,7 +243,13 @@ BROWSER_TOOLS_SCHEMA = [
     },
     {
         "name": "BrowserNavigate",
-        "description": "Navigate the browser to a URL.",
+        "description": (
+            "Navigate the browser to a URL. Use a normal page URL a person would see. "
+            "Do NOT point it at a raw JSON/API endpoint (a /api/... or search-JSON URL "
+            "like Instagram's web/search/topsearch): that paints an unreadable data wall "
+            "in the card. To READ a site's own API, use BrowserReplayRoute, which fetches "
+            "the JSON silently without disturbing the page."
+        ),
         "input_schema": {
             "type": "object",
             "properties": {