diff --git a/setup_server.py b/setup_server.py
index e1be876..5859753 100644
--- a/setup_server.py
+++ b/setup_server.py
@@ -582,15 +582,24 @@ def _write_services():
     services_dir = str(APP_DIR / "backend/services")
 
     # (wsgi_target, workers, threads, timeout_sec, port, description)
+    #
+    # NOTA backend (1 worker + 8 threads, era 2 sync workers fino al 2026-05-04):
+    #   Le sessioni utente sono mantenute in-memory (_sessions_cache in core.py)
+    #   con persistence su file (argos_sessions.json) MA senza ricarica per
+    #   worker. Con N>1 worker, il worker B non vede sessioni create da worker
+    #   A → 401 random post-login (round-robin kernel). Workaround attuale:
+    #   1 worker con 8 threads = concorrenza intra-process via GIL release su
+    #   I/O. Sufficiente per il carico ASREM. Soluzione futura: session store
+    #   condiviso (Redis o SQLite) per scalare a N>1 worker.
     services = {
-        "backend":   ("argos_backend:app",         2, 1, 120, 8080, "Backend API"),
+        "backend":   ("argos_backend:app",         1, 8, 120, 8080, "Backend API"),
         "sync":      ("argos_sync_wsgi:app",       1, 4, 600, 8081, "Sync Daemon"),
         "ops":       ("argos_ops_wsgi:app",        1, 4, 600, 8082, "Ops Daemon"),
         "analytics": ("argos_analytics_wsgi:app",  1, 4, 300, 8083, "Analytics Daemon"),
     }
 
     for svc, (wsgi, workers, threads, timeout, port, desc) in services.items():
-        # Solo i servizi con threads>1 usano gthread; backend e' sync con 2 worker.
+        # Tutti i servizi ora usano gthread (threads>=4); workers=1 obbligatorio.
         threads_line = f"    --threads {threads} \\\n" if threads > 1 else ""
         unit = f"""[Unit]
 Description=ARGOS {desc} (Gunicorn)