diff --git a/src/roadrunner_acceptor.erl b/src/roadrunner_acceptor.erl
index c5af7ad..00c2d6b 100644
--- a/src/roadrunner_acceptor.erl
+++ b/src/roadrunner_acceptor.erl
@@ -5,12 +5,21 @@
 %% and hands each accepted connection off to a `roadrunner_conn` worker.
 %%
 %% Spawn-linked to the owning `roadrunner_listener`: a listener stop closes
-%% the listen socket, the acceptor's `accept/1` returns `{error, _}`,
-%% and the acceptor exits cleanly. Unrelated acceptor crashes propagate
+%% the listen socket, the acceptor's `accept/1` returns `{error, closed}`,
+%% and the acceptor exits cleanly. A transient accept error instead
+%% (`emfile`/`enfile`/`system_limit` descriptor exhaustion when `max_clients`
+%% sits above the OS `ulimit -n`, or a connection aborted before accept
+%% completed) is reported via telemetry and retried after a short back-off,
+%% so the pool never silently drains. Unrelated acceptor crashes propagate
 %% back via the link, taking the listener down for supervisor restart.
 %% Connection workers are spawned **without** a link so that a crash
 %% in one connection does not bring down the acceptor.
 
+%% Back-off between retries after a transient accept error. Bounds the
+%% retry/telemetry rate and gives the box a moment to reclaim descriptors
+%% before the next `accept/1`.
+-define(ACCEPT_ERROR_BACKOFF_MS, 100).
+
 -export([start_link/3]).
 
 -doc """
@@ -37,10 +46,25 @@ loop(LSocket, ProtoOpts) ->
         {ok, Socket} ->
             handle_accepted(Socket, ProtoOpts),
             loop(LSocket, ProtoOpts);
-        {error, _} ->
-            %% Listen socket was closed (or another transport error) —
-            %% terminate cleanly; the linked listener will tear us down.
-            ok
+        {error, closed} ->
+            %% Listen socket was closed — the listener is stopping. Exit
+            %% cleanly; the linked listener tears the rest of the pool down.
+            ok;
+        {error, Reason} ->
+            %% Any other accept error is transient: descriptor exhaustion
+            %% (`emfile`/`enfile`/`system_limit`) when `max_clients` sits
+            %% above the OS `ulimit -n`, or a connection aborted before
+            %% accept completed. Surface it and back off, then keep
+            %% accepting — exiting here would silently drain the pool and
+            %% leave the listener permanently deaf with no diagnostic.
+            ok = roadrunner_telemetry:listener_accept_error(#{
+                listener_name => maps:get(listener_name, ProtoOpts, undefined),
+                reason => Reason
+            }),
+            receive
+            after ?ACCEPT_ERROR_BACKOFF_MS -> ok
+            end,
+            loop(LSocket, ProtoOpts)
     end.
 
 -spec handle_accepted(roadrunner_transport:socket(), roadrunner_conn:proto_opts()) -> ok.
diff --git a/src/roadrunner_telemetry.erl b/src/roadrunner_telemetry.erl
index b54a771..b71b490 100644
--- a/src/roadrunner_telemetry.erl
+++ b/src/roadrunner_telemetry.erl
@@ -76,6 +76,16 @@
 %%   - **Measurements:** `system_time`.
 %%   - **Metadata:** `listener_name`, `reason` (`max_clients`).
 %%
+%% - `[roadrunner, listener, accept_error]` — fired when an acceptor's
+%%   `accept/1` returns a transient error: file-descriptor exhaustion
+%%   (`emfile`/`enfile`/`system_limit`, usually `max_clients` above the OS
+%%   `ulimit -n`) or a connection aborted before accept completed. The
+%%   acceptor backs off and keeps accepting instead of exiting, so this is
+%%   the signal that the box is out of descriptors — raise `ulimit -n`.
+%%
+%%   - **Measurements:** `system_time`.
+%%   - **Metadata:** `listener_name`, `reason` (the accept error).
+%%
 %% - `[roadrunner, request, throttled]` — fired before any handler runs when a
 %%   request is refused at a listener limit, for one of two reasons:
 %%     - `max_concurrent_requests`: an HTTP/2 or HTTP/3 stream over the
@@ -169,6 +179,7 @@
     listener_accept/1,
     listener_conn_close/2,
     listener_conn_rejected/1,
+    listener_accept_error/1,
     request_rejected/1,
     request_throttled/1,
     slots_reconciled/1,
@@ -310,6 +321,25 @@ listener_conn_rejected(Metadata) ->
     ),
     ok.
 
+-doc """
+Emit `[roadrunner, listener, accept_error]` when an acceptor's `accept/1`
+fails with a transient error — file-descriptor exhaustion
+(`emfile`/`enfile`/`system_limit`, typically `max_clients` sitting above
+the OS `ulimit -n`) or a connection aborted before accept completed. The
+acceptor reports it here, backs off, and keeps accepting rather than
+exiting, so this event is the signal that the box is out of descriptors.
+`Metadata` should include `listener_name` and `reason` (the accept
+error). Carries no `peer`: there is no connection to name.
+""".
+-spec listener_accept_error(map()) -> ok.
+listener_accept_error(Metadata) ->
+    telemetry:execute(
+        [roadrunner, listener, accept_error],
+        #{system_time => erlang:system_time()},
+        Metadata
+    ),
+    ok.
+
 -doc """
 Emit `[roadrunner, request, rejected]` when a request is dropped at the
 parser/limit layer before any handler runs (malformed request line,
diff --git a/test/roadrunner_acceptor_tests.erl b/test/roadrunner_acceptor_tests.erl
index 35e88b6..eb33fc8 100644
--- a/test/roadrunner_acceptor_tests.erl
+++ b/test/roadrunner_acceptor_tests.erl
@@ -105,6 +105,47 @@ conn_process_carries_listener_name_and_peer_label_test_() ->
             end}
         end}.
 
+acceptor_retries_on_transient_accept_error_test() ->
+    %% Accepting on a connected (non-listen) socket fails with a transient
+    %% error (einval here, standing in for emfile descriptor exhaustion). The
+    %% acceptor must report it via telemetry and keep looping, NOT exit and
+    %% leave the listener silently deaf. Closing the socket then yields
+    %% {error, closed}, which ends the loop cleanly.
+    {ok, LSock} = gen_tcp:listen(0, [binary, {active, false}]),
+    {ok, Port} = inet:port(LSock),
+    {ok, CSock} = gen_tcp:connect({127, 0, 0, 1}, Port, [binary, {active, false}], 1000),
+    {ok, _} = application:ensure_all_started(telemetry),
+    Self = self(),
+    HandlerId = {?MODULE, make_ref()},
+    ok = telemetry:attach(
+        HandlerId,
+        [roadrunner, listener, accept_error],
+        fun(_Event, _Measure, Meta, _Cfg) -> Self ! {accept_error, Meta} end,
+        undefined
+    ),
+    {ok, Pid} = roadrunner_acceptor:start_link(
+        {gen_tcp, CSock}, #{listener_name => acceptor_test_accept_error}, 1
+    ),
+    receive
+        {accept_error, Meta} ->
+            ?assertEqual(acceptor_test_accept_error, maps:get(listener_name, Meta)),
+            ?assertNotEqual(closed, maps:get(reason, Meta))
+    after 2000 ->
+        error(no_accept_error_telemetry)
+    end,
+    %% Survived the transient error.
+    ?assert(is_process_alive(Pid)),
+    %% Closing the socket turns the next accept into {error, closed} → exit.
+    MRef = erlang:monitor(process, Pid),
+    ok = gen_tcp:close(CSock),
+    receive
+        {'DOWN', MRef, process, Pid, normal} -> ok
+    after 2000 ->
+        error(acceptor_did_not_stop)
+    end,
+    ok = telemetry:detach(HandlerId),
+    ok = gen_tcp:close(LSock).
+
 %% --- helpers ---
 
 %% Poll until we see a refined `{roadrunner_conn, Name, Peer}` label or run