diff --git a/agent/auxiliary_client.py b/agent/auxiliary_client.py
index 22bf9c165c..5a33d25cc7 100644
--- a/agent/auxiliary_client.py
+++ b/agent/auxiliary_client.py
@@ -7601,6 +7601,77 @@ def _get_task_extra_body(task: str) -> Dict[str, Any]:
     return result
 
 
+# ---------------------------------------------------------------------------
+# Per-task concurrency limiting (#23324)
+# ---------------------------------------------------------------------------
+# Background auxiliary work (title generation, context compression, etc.) can
+# spawn unbounded concurrent LLM calls when many sessions are active. During
+# provider incidents each call also retries / fans out across the fallback
+# chain, multiplying request volume on already-degraded endpoints. A per-task
+# semaphore caps in-flight calls so retry amplification stays bounded.
+
+_aux_sync_semaphores: Dict[str, Tuple[int, threading.BoundedSemaphore]] = {}
+_aux_async_semaphores: Dict[Tuple[str, int], Tuple[int, Any]] = {}
+_aux_sem_lock = threading.Lock()
+
+
+def _get_task_max_concurrency(task: Optional[str]) -> Optional[int]:
+    """Return ``auxiliary.<task>.max_concurrency`` as a positive int, or None."""
+    if not task or task == "vision":
+        # Vision already uses this key for its encode/resize CPU worker pool;
+        # its LLM calls deliberately remain concurrent.
+        return None
+    raw = _get_auxiliary_task_config(task).get("max_concurrency")
+    if raw is None:
+        return None
+    try:
+        value = int(raw)
+    except (TypeError, ValueError):
+        return None
+    return value if value > 0 else None
+
+
+def _acquire_sync_aux_semaphore(task: Optional[str]) -> Optional[threading.BoundedSemaphore]:
+    """Get a per-task sync semaphore, rebuilding it after a config change."""
+    limit = _get_task_max_concurrency(task)
+    if limit is None:
+        return None
+    with _aux_sem_lock:
+        entry = _aux_sync_semaphores.get(task)
+        if entry is None or entry[0] != limit:
+            semaphore = threading.BoundedSemaphore(limit)
+            _aux_sync_semaphores[task] = (limit, semaphore)
+            return semaphore
+        return entry[1]
+
+
+def _acquire_async_aux_semaphore(task: Optional[str]):
+    """Get a per-task, per-event-loop async semaphore after config lookup."""
+    limit = _get_task_max_concurrency(task)
+    if limit is None:
+        return None
+    import asyncio
+    try:
+        loop = asyncio.get_running_loop()
+    except RuntimeError:
+        return None
+    key = (task, id(loop))
+    with _aux_sem_lock:
+        entry = _aux_async_semaphores.get(key)
+        if entry is None or entry[0] != limit:
+            semaphore = asyncio.Semaphore(limit)
+            _aux_async_semaphores[key] = (limit, semaphore)
+            return semaphore
+        return entry[1]
+
+
+def _reset_aux_semaphores() -> None:
+    """Drop cached semaphores (test helper)."""
+    with _aux_sem_lock:
+        _aux_sync_semaphores.clear()
+        _aux_async_semaphores.clear()
+
+
 # ---------------------------------------------------------------------------
 # Anthropic-compatible endpoint detection + image block conversion
 # ---------------------------------------------------------------------------
@@ -8455,6 +8526,73 @@ def call_llm(
     api_mode: str = None,
     stream: bool = False,
     stream_options: dict = None,
+) -> Any:
+    """Run an auxiliary LLM request, applying the configured task limit."""
+    semaphore = _acquire_sync_aux_semaphore(task)
+    if semaphore is not None:
+        semaphore.acquire()
+    try:
+        response = _call_llm_impl(
+            task=task,
+            provider=provider,
+            model=model,
+            base_url=base_url,
+            api_key=api_key,
+            main_runtime=main_runtime,
+            messages=messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+            tools=tools,
+            timeout=timeout,
+            extra_body=extra_body,
+            reasoning_config=reasoning_config,
+            api_mode=api_mode,
+            stream=stream,
+            stream_options=stream_options,
+        )
+        if stream and semaphore is not None:
+            stream_semaphore = semaphore
+            semaphore = None
+            return _release_sync_semaphore_after_stream(response, stream_semaphore)
+        return response
+    finally:
+        if semaphore is not None:
+            semaphore.release()
+
+
+def _release_sync_semaphore_after_stream(
+    stream: Any, semaphore: threading.BoundedSemaphore,
+):
+    """Release a permit only after a streaming response is consumed or closed."""
+    try:
+        yield from stream
+    finally:
+        try:
+            close = getattr(stream, "close", None)
+            if callable(close):
+                close()
+        finally:
+            semaphore.release()
+
+
+def _call_llm_impl(
+    task: str = None,
+    *,
+    provider: str = None,
+    model: str = None,
+    base_url: str = None,
+    api_key: str = None,
+    main_runtime: Optional[Dict[str, Any]] = None,
+    messages: list,
+    temperature: Optional[float] = None,
+    max_tokens: int = None,
+    tools: list = None,
+    timeout: float = None,
+    extra_body: dict = None,
+    reasoning_config: Optional[dict] = None,
+    api_mode: str = None,
+    stream: bool = False,
+    stream_options: dict = None,
 ) -> Any:
     """Centralized synchronous LLM call.
 
@@ -9218,6 +9356,47 @@ async def async_call_llm(
     timeout: float = None,
     extra_body: dict = None,
     reasoning_config: Optional[dict] = None,
+) -> Any:
+    """Run an asynchronous auxiliary LLM request under the configured limit."""
+    semaphore = _acquire_async_aux_semaphore(task)
+    if semaphore is not None:
+        await semaphore.acquire()
+    try:
+        return await _async_call_llm_impl(
+            task=task,
+            provider=provider,
+            model=model,
+            base_url=base_url,
+            api_key=api_key,
+            main_runtime=main_runtime,
+            messages=messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+            tools=tools,
+            timeout=timeout,
+            extra_body=extra_body,
+            reasoning_config=reasoning_config,
+        )
+    finally:
+        if semaphore is not None:
+            semaphore.release()
+
+
+async def _async_call_llm_impl(
+    task: str = None,
+    *,
+    provider: str = None,
+    model: str = None,
+    base_url: str = None,
+    api_key: str = None,
+    main_runtime: Optional[Dict[str, Any]] = None,
+    messages: list,
+    temperature: Optional[float] = None,
+    max_tokens: int = None,
+    tools: list = None,
+    timeout: float = None,
+    extra_body: dict = None,
+    reasoning_config: Optional[dict] = None,
 ) -> Any:
     """Centralized asynchronous LLM call.
 
diff --git a/cli-config.yaml.example b/cli-config.yaml.example
index 04ccfb6da4..6b7a4c327a 100644
--- a/cli-config.yaml.example
+++ b/cli-config.yaml.example
@@ -652,6 +652,25 @@ prompt_caching:
 #                           # extra_body:
 #                           #   chat_template_kwargs:
 #                           #     enable_thinking: false
+#
+#   # Auto-generated short session titles after the first exchange.
+#   # Each active Discord/Telegram channel can spawn a background title
+#   # call.  Cap concurrency to keep retries during provider incidents
+#   # from amplifying the request burst.  Leave unset for legacy behavior
+#   # (unlimited).
+#   title_generation:
+#     provider: "auto"
+#     model: ""
+#     # max_concurrency: 2    # Optional: cap simultaneous title calls
+#
+#   # Context compression — summarizes long sessions to shrink the prompt.
+#   # Heavy and often hits the slowest provider chain.  Setting a small
+#   # cap prevents many sessions from compressing simultaneously during
+#   # provider degradation.  Leave unset for legacy behavior (unlimited).
+#   compression:
+#     provider: "auto"
+#     model: ""
+#     # max_concurrency: 2    # Optional: cap simultaneous compression calls
 
 # =============================================================================
 # Persistent Memory
diff --git a/tests/agent/test_auxiliary_concurrency.py b/tests/agent/test_auxiliary_concurrency.py
new file mode 100644
index 0000000000..27fccd97ec
--- /dev/null
+++ b/tests/agent/test_auxiliary_concurrency.py
@@ -0,0 +1,403 @@
+"""Tests for per-task concurrency limiting on auxiliary LLM calls (#23324)."""
+
+import asyncio
+import threading
+import time
+from unittest.mock import MagicMock, AsyncMock, patch
+
+import pytest
+
+from agent.auxiliary_client import (
+    call_llm,
+    async_call_llm,
+    _acquire_sync_aux_semaphore,
+    _acquire_async_aux_semaphore,
+    _get_task_max_concurrency,
+    _reset_aux_semaphores,
+)
+
+
+@pytest.fixture(autouse=True)
+def _clean_semaphore_cache():
+    _reset_aux_semaphores()
+    yield
+    _reset_aux_semaphores()
+
+
+class TestGetTaskMaxConcurrency:
+    def test_returns_none_for_missing_task(self):
+        assert _get_task_max_concurrency(None) is None
+        assert _get_task_max_concurrency("") is None
+
+    def test_returns_none_when_unset(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config", return_value={}
+        ):
+            assert _get_task_max_concurrency("title_generation") is None
+
+    def test_does_not_reuse_vision_cpu_limit_for_llm_calls(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value={"max_concurrency": 1},
+        ):
+            assert _get_task_max_concurrency("vision") is None
+
+    def test_returns_int_when_configured(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value={"max_concurrency": 3},
+        ):
+            assert _get_task_max_concurrency("compression") == 3
+
+    def test_returns_none_for_non_numeric(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value={"max_concurrency": "not-a-number"},
+        ):
+            assert _get_task_max_concurrency("compression") is None
+
+    def test_returns_none_for_zero_or_negative(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value={"max_concurrency": 0},
+        ):
+            assert _get_task_max_concurrency("compression") is None
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value={"max_concurrency": -2},
+        ):
+            assert _get_task_max_concurrency("compression") is None
+
+
+class TestSemaphoreCache:
+    def test_sync_returns_none_when_unset(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config", return_value={}
+        ):
+            assert _acquire_sync_aux_semaphore("title_generation") is None
+
+    def test_sync_reuses_semaphore_for_same_limit(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value={"max_concurrency": 2},
+        ):
+            sem1 = _acquire_sync_aux_semaphore("compression")
+            sem2 = _acquire_sync_aux_semaphore("compression")
+            assert sem1 is sem2
+
+    def test_sync_rebuilds_when_limit_changes(self):
+        cfg = {"max_concurrency": 2}
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value=cfg,
+        ):
+            sem1 = _acquire_sync_aux_semaphore("compression")
+            cfg["max_concurrency"] = 5
+            sem2 = _acquire_sync_aux_semaphore("compression")
+            assert sem1 is not sem2
+
+    @pytest.mark.asyncio
+    async def test_async_reuses_semaphore_within_same_loop(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value={"max_concurrency": 2},
+        ):
+            sem1 = _acquire_async_aux_semaphore("compression")
+            sem2 = _acquire_async_aux_semaphore("compression")
+            assert sem1 is sem2
+
+    def test_async_returns_none_with_no_running_loop(self):
+        with patch(
+            "agent.auxiliary_client._get_auxiliary_task_config",
+            return_value={"max_concurrency": 2},
+        ):
+            # Called outside an asyncio loop — should bail rather than crash.
+            assert _acquire_async_aux_semaphore("compression") is None
+
+
+class TestSyncCallEnforcesLimit:
+    def test_call_llm_caps_concurrent_inflight(self):
+        limit = 2
+        n_callers = 6
+
+        active = 0
+        max_active = 0
+        lock = threading.Lock()
+
+        def fake_create(**kwargs):
+            nonlocal active, max_active
+            with lock:
+                active += 1
+                if active > max_active:
+                    max_active = active
+            try:
+                time.sleep(0.05)
+            finally:
+                with lock:
+                    active -= 1
+            return MagicMock()
+
+        client = MagicMock()
+        client.base_url = "https://example.test/v1"
+        client.chat.completions.create.side_effect = fake_create
+
+        with (
+            patch(
+                "agent.auxiliary_client._resolve_task_provider_model",
+                return_value=("openrouter", "test-model", None, None, None),
+            ),
+            patch(
+                "agent.auxiliary_client._get_cached_client",
+                return_value=(client, "test-model"),
+            ),
+            patch(
+                "agent.auxiliary_client._validate_llm_response",
+                side_effect=lambda resp, _task, **_kwargs: resp,
+            ),
+            patch(
+                "agent.auxiliary_client._get_auxiliary_task_config",
+                return_value={"max_concurrency": limit},
+            ),
+        ):
+            threads = [
+                threading.Thread(
+                    target=lambda: call_llm(
+                        task="title_generation",
+                        messages=[{"role": "user", "content": "hi"}],
+                    )
+                )
+                for _ in range(n_callers)
+            ]
+            for t in threads:
+                t.start()
+            for t in threads:
+                t.join(timeout=5)
+
+        assert max_active <= limit, f"observed {max_active} > limit {limit}"
+        assert client.chat.completions.create.call_count == n_callers
+
+    def test_call_llm_unlimited_when_not_configured(self):
+        client = MagicMock()
+        client.base_url = "https://example.test/v1"
+        client.chat.completions.create.return_value = MagicMock()
+
+        with (
+            patch(
+                "agent.auxiliary_client._resolve_task_provider_model",
+                return_value=("openrouter", "test-model", None, None, None),
+            ),
+            patch(
+                "agent.auxiliary_client._get_cached_client",
+                return_value=(client, "test-model"),
+            ),
+            patch(
+                "agent.auxiliary_client._validate_llm_response",
+                side_effect=lambda resp, _task, **_kwargs: resp,
+            ),
+            patch(
+                "agent.auxiliary_client._get_auxiliary_task_config",
+                return_value={},
+            ),
+        ):
+            # With no max_concurrency in config, no semaphore is acquired.
+            call_llm(
+                task="title_generation",
+                messages=[{"role": "user", "content": "hi"}],
+            )
+
+        assert client.chat.completions.create.call_count == 1
+
+    def test_semaphore_released_on_exception(self):
+        """Errors inside call_llm must release the semaphore so the next call proceeds."""
+        client = MagicMock()
+        client.base_url = "https://example.test/v1"
+        client.chat.completions.create.side_effect = RuntimeError("boom")
+
+        with (
+            patch(
+                "agent.auxiliary_client._resolve_task_provider_model",
+                return_value=("openrouter", "test-model", None, None, None),
+            ),
+            patch(
+                "agent.auxiliary_client._get_cached_client",
+                return_value=(client, "test-model"),
+            ),
+            patch(
+                "agent.auxiliary_client._validate_llm_response",
+                side_effect=lambda resp, _task, **_kwargs: resp,
+            ),
+            patch(
+                "agent.auxiliary_client._get_auxiliary_task_config",
+                return_value={"max_concurrency": 1},
+            ),
+        ):
+            for _ in range(3):
+                with pytest.raises(RuntimeError, match="boom"):
+                    call_llm(
+                        task="title_generation",
+                        messages=[{"role": "user", "content": "hi"}],
+                    )
+
+    def test_stream_holds_permit_until_consumed_and_preserves_options(self):
+        client = MagicMock()
+        client.base_url = "https://example.test/v1"
+        client.chat.completions.create.side_effect = [iter(["chunk"]), MagicMock()]
+        second_call_started = threading.Event()
+
+        def make_second_call():
+            second_call_started.set()
+            call_llm(
+                task="compression",
+                messages=[{"role": "user", "content": "second"}],
+            )
+
+        with (
+            patch(
+                "agent.auxiliary_client._resolve_task_provider_model",
+                return_value=("openrouter", "test-model", None, None, None),
+            ),
+            patch(
+                "agent.auxiliary_client._get_cached_client",
+                return_value=(client, "test-model"),
+            ),
+            patch(
+                "agent.auxiliary_client._validate_llm_response",
+                side_effect=lambda response, _task, **_kwargs: response,
+            ),
+            patch(
+                "agent.auxiliary_client._get_auxiliary_task_config",
+                return_value={"max_concurrency": 1},
+            ),
+        ):
+            stream = call_llm(
+                task="compression",
+                messages=[{"role": "user", "content": "first"}],
+                stream=True,
+                stream_options={"include_usage": True},
+            )
+            thread = threading.Thread(target=make_second_call)
+            thread.start()
+            assert second_call_started.wait(timeout=1)
+            time.sleep(0.05)
+            assert client.chat.completions.create.call_count == 1
+            assert list(stream) == ["chunk"]
+            thread.join(timeout=1)
+
+        assert not thread.is_alive()
+        assert client.chat.completions.create.call_count == 2
+        assert client.chat.completions.create.call_args_list[0].kwargs["stream"] is True
+        assert client.chat.completions.create.call_args_list[0].kwargs["stream_options"] == {
+            "include_usage": True
+        }
+
+    def test_api_mode_is_forwarded_to_client_resolution(self):
+        client = MagicMock()
+        client.base_url = "https://example.test/v1"
+        client.chat.completions.create.return_value = MagicMock()
+
+        with (
+            patch(
+                "agent.auxiliary_client._resolve_task_provider_model",
+                return_value=("openrouter", "test-model", None, None, None),
+            ),
+            patch(
+                "agent.auxiliary_client._get_cached_client",
+                return_value=(client, "test-model"),
+            ) as get_client,
+            patch(
+                "agent.auxiliary_client._validate_llm_response",
+                side_effect=lambda response, _task, **_kwargs: response,
+            ),
+        ):
+            call_llm(
+                task="title_generation",
+                messages=[{"role": "user", "content": "hi"}],
+                api_mode="codex_responses",
+            )
+
+        assert get_client.call_args.kwargs["api_mode"] == "codex_responses"
+
+
+class TestAsyncCallEnforcesLimit:
+    @pytest.mark.asyncio
+    async def test_async_call_llm_caps_concurrent_inflight(self):
+        limit = 2
+        n_callers = 6
+
+        active = 0
+        max_active = 0
+
+        async def fake_create(**kwargs):
+            nonlocal active, max_active
+            active += 1
+            if active > max_active:
+                max_active = active
+            try:
+                await asyncio.sleep(0.05)
+            finally:
+                active -= 1
+            return MagicMock()
+
+        client = MagicMock()
+        client.base_url = "https://example.test/v1"
+        client.chat.completions.create = AsyncMock(side_effect=fake_create)
+
+        with (
+            patch(
+                "agent.auxiliary_client._resolve_task_provider_model",
+                return_value=("openrouter", "test-model", None, None, None),
+            ),
+            patch(
+                "agent.auxiliary_client._get_cached_client",
+                return_value=(client, "test-model"),
+            ),
+            patch(
+                "agent.auxiliary_client._validate_llm_response",
+                side_effect=lambda resp, _task, **_kwargs: resp,
+            ),
+            patch(
+                "agent.auxiliary_client._get_auxiliary_task_config",
+                return_value={"max_concurrency": limit},
+            ),
+        ):
+            await asyncio.gather(*[
+                async_call_llm(
+                    task="compression",
+                    messages=[{"role": "user", "content": "hi"}],
+                )
+                for _ in range(n_callers)
+            ])
+
+        assert max_active <= limit, f"observed {max_active} > limit {limit}"
+        assert client.chat.completions.create.await_count == n_callers
+
+    @pytest.mark.asyncio
+    async def test_async_semaphore_released_on_exception(self):
+        client = MagicMock()
+        client.base_url = "https://example.test/v1"
+        client.chat.completions.create = AsyncMock(side_effect=RuntimeError("boom"))
+
+        with (
+            patch(
+                "agent.auxiliary_client._resolve_task_provider_model",
+                return_value=("openrouter", "test-model", None, None, None),
+            ),
+            patch(
+                "agent.auxiliary_client._get_cached_client",
+                return_value=(client, "test-model"),
+            ),
+            patch(
+                "agent.auxiliary_client._validate_llm_response",
+                side_effect=lambda resp, _task, **_kwargs: resp,
+            ),
+            patch(
+                "agent.auxiliary_client._get_auxiliary_task_config",
+                return_value={"max_concurrency": 1},
+            ),
+        ):
+            for _ in range(3):
+                with pytest.raises(RuntimeError, match="boom"):
+                    await async_call_llm(
+                        task="compression",
+                        messages=[{"role": "user", "content": "hi"}],
+                    )
diff --git a/website/docs/user-guide/configuration.md b/website/docs/user-guide/configuration.md
index b1fddd57c3..cfe85dbd55 100644
--- a/website/docs/user-guide/configuration.md
+++ b/website/docs/user-guide/configuration.md
@@ -1189,6 +1189,8 @@ auxiliary:
     #     model: google/gemini-2.5-flash
     #     base_url: ""
     #     api_key: ""
+    # max_concurrency: 2       # Optional: cap simultaneous compression LLM calls so
+                               # multiple sessions don't pile retries on a degraded provider
 
   # Auto-generated session titles. Empty language follows the conversation;
   # set e.g. "English" or "Japanese" to pin titles to one language.
@@ -1217,6 +1219,15 @@ auxiliary:
     api_key: ""
     timeout: 30
 
+  # Auto-generated short session titles after the first exchange
+  title_generation:
+    provider: "auto"
+    model: ""
+    base_url: ""
+    api_key: ""
+    timeout: 30
+    # max_concurrency: 2       # Optional: cap simultaneous title-generation calls
+
   # Kanban triage specifier — `hermes kanban specify <id>` (or the
   # dashboard's ✨ Specify button on Triage-column cards) uses this
   # slot to expand a one-liner into a concrete spec and promote the
@@ -1266,6 +1277,25 @@ Each entry supports the same three knobs as any auxiliary task config:
 
 `fallback_chain` is available on any auxiliary task — `compression`, `vision`, `web_extract`, `approval`, `skills_hub`, `mcp`, etc.
 
+### Limiting auxiliary concurrency
+
+`max_concurrency` caps in-flight LLM calls for auxiliary tasks such as `compression` and `title_generation` across the whole process. `auxiliary.vision.max_concurrency` is excluded: it already controls only vision's CPU-bound image encode/resize workers, not LLM requests. This is most useful when:
+
+- Many sessions can spawn background work simultaneously (Discord/Telegram channels, multiple terminals)
+- Your provider is rate-limited or going through an incident and retries would amplify the burst
+
+The default is unlimited. A typical safety cap is `2`:
+
+```yaml
+auxiliary:
+  title_generation:
+    max_concurrency: 2
+  compression:
+    max_concurrency: 2
+```
+
+The semaphore wraps the entire call including retries and fallbacks, so a single slow call counts only once toward the limit.
+
 ### OpenRouter routing & Pareto Code for auxiliary tasks
 
 When an auxiliary task resolves to OpenRouter (either explicitly or via `provider: "main"` while your main agent is on OpenRouter), the main agent's `provider_routing` and `openrouter.min_coding_score` settings **do not propagate** — by design, each auxiliary task is independent. To set OpenRouter provider preferences or use the [Pareto Code router](/integrations/providers#openrouter-pareto-code-router) for a specific aux task, set them per-task via `extra_body`: