diff --git a/examples/hermes_sentinel/hermes_bridge.py b/examples/hermes_sentinel/hermes_bridge.py
index 01df109..5cdd90a 100755
--- a/examples/hermes_sentinel/hermes_bridge.py
+++ b/examples/hermes_sentinel/hermes_bridge.py
@@ -64,10 +64,36 @@
 
 from __future__ import annotations
 
+import json
 import os
 import sys
+import uuid
 from pathlib import Path
 
+# ─── AX_GATEWAY_EVENT protocol ────────────────────────────────────────────
+# Lines prefixed with `AX_GATEWAY_EVENT ` on stdout are parsed by the Gateway
+# and forwarded as platform `agent_processing` / tool-call events so the UI
+# sees per-mention phase text while the agent is still working. Unprefixed
+# stdout lines accumulate into the final reply body.
+EVENT_PREFIX = "AX_GATEWAY_EVENT "
+
+# Preserve the *original* stdout before we suppress hermes's internal prints.
+# AX_GATEWAY_EVENT lines and the final reply must reach Gateway, but hermes's
+# chatter ("deliberating...", tool previews, etc.) must not — otherwise the
+# Gateway captures it all as the reply body. `_real_stdout` is the escape hatch
+# for our own prints.
+_real_stdout = sys.stdout
+
+
+def _emit_event(payload: dict) -> None:
+    _real_stdout.write(f"{EVENT_PREFIX}{json.dumps(payload, sort_keys=True)}\n")
+    _real_stdout.flush()
+
+
+def _print_reply(text: str) -> None:
+    _real_stdout.write(text + "\n")
+    _real_stdout.flush()
+
 # ─── Resolve hermes-agent location ─────────────────────────────────────────
 HERMES_REPO = Path(
     os.environ.get("HERMES_REPO_PATH", str(Path.home() / "hermes-agent"))
@@ -200,6 +226,77 @@ def main() -> int:
     # Change to workdir so relative file tool paths behave predictably.
     os.chdir(workdir)
 
+    # ─── Phase events for Gateway/UI ───────────────────────────────────────
+    # `tool_progress_callback` is an AIAgent constructor param (see
+    # hermes_agent/run_agent.py:446). It fires before each tool call with
+    # (name, args_preview, args_dict). We translate those into
+    # AX_GATEWAY_EVENT tool_start/tool_result pairs so the UI chip
+    # reflects what Hermes is actually doing.
+    #
+    # `activity` is the human-readable chip text (<40 chars). Extracting
+    # a salient arg per tool makes the bubble read
+    #   "@X: $ pwd"  or  "@X: reading AGENTS.md"
+    # instead of the generic "@X is using terminal…".
+
+    def _tool_activity(tool_name: str, args: dict | None) -> str:
+        args = args if isinstance(args, dict) else {}
+        name = (tool_name or "").lower()
+        # Normalize synonyms across hermes's tool namespace
+        if name in {"terminal", "bash", "shell", "run_command"}:
+            cmd = args.get("command") or args.get("cmd") or ""
+            cmd = str(cmd).splitlines()[0].strip() if cmd else ""
+            return f"$ {cmd}"[:80] if cmd else "running shell"
+        if name in {"read_file", "read", "view"}:
+            path = args.get("path") or args.get("file") or args.get("filename") or ""
+            return f"reading {path}"[:80] if path else "reading file"
+        if name in {"write_file", "write", "create_file"}:
+            path = args.get("path") or args.get("file") or args.get("filename") or ""
+            return f"writing {path}"[:80] if path else "writing file"
+        if name in {"edit_file", "edit", "str_replace"}:
+            path = args.get("path") or args.get("file") or ""
+            return f"editing {path}"[:80] if path else "editing file"
+        if name in {"search", "grep", "find"}:
+            pattern = args.get("pattern") or args.get("query") or args.get("q") or ""
+            return f"searching {pattern}"[:80] if pattern else "searching"
+        if name in {"web_search", "search_web"}:
+            q = args.get("query") or args.get("q") or ""
+            return f"web search: {q}"[:80] if q else "web search"
+        if name in {"fetch", "http_get", "url"}:
+            url = args.get("url") or args.get("target") or ""
+            return f"fetching {url}"[:80] if url else "fetching"
+        # Generic fallback: surface the first scalar arg value
+        for k, v in args.items():
+            if isinstance(v, (str, int, float, bool)):
+                return f"{tool_name}: {str(v)}"[:80]
+        return f"using {tool_name}"[:80]
+
+    def _on_tool_progress(tool_name: str, args_preview: str, args_dict=None):
+        tool_call_id = f"hermes-{uuid.uuid4()}"
+        args = args_dict if isinstance(args_dict, dict) else {}
+        activity = _tool_activity(tool_name, args)
+        try:
+            _emit_event({
+                "kind": "tool_start",
+                "tool_name": tool_name,
+                "tool_action": tool_name,
+                "tool_call_id": tool_call_id,
+                "status": "tool_call",
+                "arguments": args,
+                "activity": activity,
+                "message": activity,
+            })
+            _emit_event({
+                "kind": "tool_result",
+                "tool_name": tool_name,
+                "tool_action": tool_name,
+                "tool_call_id": tool_call_id,
+                "status": "tool_complete",
+                "activity": activity,
+                "message": activity,
+            })
+        except Exception:
+            pass  # never let event emission break the agent run
+
     agent = AIAgent(
         base_url=cfg["base_url"],
         api_key=cfg["api_key"],
@@ -215,21 +312,51 @@ def main() -> int:
             "web", "browser", "image_generation", "tts", "vision",
             "cronjob", "rl_training", "homeassistant",
         ],
+        tool_progress_callback=_on_tool_progress,
     )
 
+    _emit_event({"kind": "status", "status": "started", "message": "Agent planning"})
+    _emit_event({"kind": "status", "status": "thinking", "message": "Thinking"})
+
     # ─── Run a single conversation turn ────────────────────────────────────
-    result = agent.run_conversation(
-        user_message=content,
-        system_message=system_prompt,
-    )
+    # Hermes's AIAgent prints progress chatter ("deliberating...", tool
+    # previews, "🔧 ..." etc.) to stdout even with quiet_mode=True.  If we
+    # don't suppress it, Gateway captures all of it as the reply body.
+    # Redirect sys.stdout/sys.stderr to devnull during the run; our own
+    # AX_GATEWAY_EVENT emissions write to `_real_stdout` directly and
+    # still reach Gateway.
+    saved_stdout = sys.stdout
+    saved_stderr = sys.stderr
+    devnull = open(os.devnull, "w")
+    sys.stdout = devnull
+    sys.stderr = devnull
+    try:
+        try:
+            result = agent.run_conversation(
+                user_message=content,
+                system_message=system_prompt,
+            )
+        except Exception as run_err:
+            sys.stdout = saved_stdout
+            sys.stderr = saved_stderr
+            _emit_event({"kind": "status", "status": "error", "message": f"Agent error: {run_err}"[:200]})
+            print(f"Hermes bridge failed: {run_err}", file=saved_stderr)
+            return 1
+    finally:
+        sys.stdout = saved_stdout
+        sys.stderr = saved_stderr
+        devnull.close()
 
     final_text = result.get("final_response", "").strip()
     if not final_text:
+        _emit_event({"kind": "status", "status": "error", "message": "Agent produced no output"})
         print("(agent produced no output)", file=sys.stderr)
         return 1
 
-    # ax listen captures stdout → posts to aX as the agent's reply.
-    print(final_text)
+    _emit_event({"kind": "status", "status": "completed", "message": "Reply ready"})
+
+    # Gateway captures unprefixed stdout lines → posts to aX as the reply.
+    _print_reply(final_text)
     return 0