diff --git a/src/agentex/lib/core/observability/event_metrics.py b/src/agentex/lib/core/observability/event_metrics.py
new file mode 100644
index 000000000..6669f2790
--- /dev/null
+++ b/src/agentex/lib/core/observability/event_metrics.py
@@ -0,0 +1,89 @@
+"""OTel metrics for event delivery (task/event_send -> live workflow).
+
+Records whether an EVENT_SEND signal actually reached a *running* Temporal
+workflow, vs. only being accepted by the ACP server. ``event/send`` is async:
+the ACP server returns a 200 ack before the signal is even attempted (see
+``base_acp_server._handle_jsonrpc`` background branch), so HTTP status is not a
+delivery signal — this counter is. Under load, tasks that have already
+completed or idle-timed-out have no workflow to receive the event, which
+Temporal surfaces as an ``RPCError`` with status ``NOT_FOUND``.
+
+The meter is no-op when the application hasn't configured a ``MeterProvider``,
+so importing this module is safe for runtimes that don't use OTel. Instruments
+are created lazily on first ``get_event_metrics()`` call so a ``MeterProvider``
+configured *after* this module is imported still binds correctly.
+
+Recording is gated on ``AGENTEX_EVENT_METRICS`` (default on) and every record
+is best-effort — it never raises into the business path.
+
+Cardinality is bounded: the only attribute is ``outcome``, drawn from a small
+fixed set (see the ``OUTCOME_*`` constants). Resource attributes
+(``service.name``, ``k8s.*``, etc.) come from the application's OTel resource
+configuration and are added to every series automatically.
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Optional
+
+from opentelemetry import metrics
+
+# Outcome label values (bounded cardinality — do NOT add task IDs etc.).
+OUTCOME_DELIVERED = "delivered"
+OUTCOME_NO_LIVE_WORKFLOW = "no_live_workflow"
+OUTCOME_ERROR = "error"
+
+
+class EventMetrics:
+    """Lazily-created OTel instruments for event-delivery telemetry."""
+
+    def __init__(self) -> None:
+        meter = metrics.get_meter("agentex.events")
+        self.delivery = meter.create_counter(
+            name="agentex.events.delivery",
+            unit="1",
+            description=(
+                "task/event_send deliveries tagged with outcome (delivered / "
+                "no_live_workflow / error). 'no_live_workflow' means the signal "
+                "had no running workflow to receive it (task already completed "
+                "or idle-timed-out). Use delivered / (delivered + "
+                "no_live_workflow) as the true event-delivery rate."
+            ),
+        )
+
+
+_event_metrics: Optional[EventMetrics] = None
+
+
+def get_event_metrics() -> EventMetrics:
+    """Return the event metrics singleton, creating it on first use."""
+    global _event_metrics
+    if _event_metrics is None:
+        _event_metrics = EventMetrics()
+    return _event_metrics
+
+
+_metrics_enabled: Optional[bool] = None
+
+
+def is_event_metrics_enabled() -> bool:
+    """Whether event-delivery metric recording is enabled (AGENTEX_EVENT_METRICS)."""
+    global _metrics_enabled
+    if _metrics_enabled is None:
+        raw = os.environ.get("AGENTEX_EVENT_METRICS", "1").strip().lower()
+        _metrics_enabled = raw not in ("0", "false", "no", "off")
+    return _metrics_enabled
+
+
+def record_event_delivery(outcome: str) -> None:
+    """Best-effort bump of the event-delivery counter for the given outcome.
+
+    ``outcome`` should be one of the ``OUTCOME_*`` constants.
+    """
+    if not is_event_metrics_enabled():
+        return
+    try:
+        get_event_metrics().delivery.add(1, {"outcome": outcome})
+    except Exception:
+        pass
diff --git a/src/agentex/lib/core/temporal/services/temporal_task_service.py b/src/agentex/lib/core/temporal/services/temporal_task_service.py
index 5f6c0c381..ac0513f74 100644
--- a/src/agentex/lib/core/temporal/services/temporal_task_service.py
+++ b/src/agentex/lib/core/temporal/services/temporal_task_service.py
@@ -3,6 +3,8 @@
 from typing import Any
 from datetime import timedelta
 
+from temporalio.service import RPCError, RPCStatusCode
+
 from agentex.types.task import Task
 from agentex.types.agent import Agent
 from agentex.types.event import Event
@@ -11,6 +13,12 @@
 from agentex.lib.core.clients.temporal.types import WorkflowState
 from agentex.lib.core.temporal.types.workflow import SignalName
 from agentex.lib.core.clients.temporal.temporal_client import TemporalClient
+from agentex.lib.core.observability.event_metrics import (
+    OUTCOME_DELIVERED,
+    OUTCOME_ERROR,
+    OUTCOME_NO_LIVE_WORKFLOW,
+    record_event_delivery,
+)
 
 
 class TemporalTaskService:
@@ -63,16 +71,36 @@ async def get_state(self, task_id: str) -> WorkflowState:
         )
 
     async def send_event(self, agent: Agent, task: Task, event: Event, request: dict | None = None) -> None:
-        return await self._temporal_client.send_signal(
-            workflow_id=task.id,
-            signal=SignalName.RECEIVE_EVENT.value,
-            payload=SendEventParams(
-                agent=agent,
-                task=task,
-                event=event,
-                request=request,
-            ).model_dump(),
-        )
+        # event/send is accepted+acked by the ACP server before the signal is
+        # attempted, so the only place we learn whether the event actually
+        # reached a *running* workflow is here. Record the true delivery outcome
+        # (see agentex.events.delivery). Behaviour is unchanged — we still raise.
+        try:
+            result = await self._temporal_client.send_signal(
+                workflow_id=task.id,
+                signal=SignalName.RECEIVE_EVENT.value,
+                payload=SendEventParams(
+                    agent=agent,
+                    task=task,
+                    event=event,
+                    request=request,
+                ).model_dump(),
+            )
+        except RPCError as e:
+            # NOT_FOUND == no running workflow to receive the event (task already
+            # completed or idle-timed-out). Distinct from unexpected errors so we
+            # can measure the "arrived too late" rate separately.
+            record_event_delivery(
+                OUTCOME_NO_LIVE_WORKFLOW
+                if e.status == RPCStatusCode.NOT_FOUND
+                else OUTCOME_ERROR
+            )
+            raise
+        except Exception:
+            record_event_delivery(OUTCOME_ERROR)
+            raise
+        record_event_delivery(OUTCOME_DELIVERED)
+        return result
 
     async def interrupt(self, agent: Agent, task: Task, request: dict | None = None) -> None:
         """Forward a task/interrupt to the running workflow as a dedicated signal.
diff --git a/src/agentex/lib/core/temporal/workers/worker.py b/src/agentex/lib/core/temporal/workers/worker.py
index 2b4958b1f..e6c1dcff3 100644
--- a/src/agentex/lib/core/temporal/workers/worker.py
+++ b/src/agentex/lib/core/temporal/workers/worker.py
@@ -151,6 +151,7 @@ def __init__(
         task_queue,
         max_workers: int = 10,
         max_concurrent_activities: int = 10,
+        max_cached_workflows: int = 1000,
         health_check_port: int | None = None,
         plugins: list = [],
         interceptors: list = [],
@@ -162,6 +163,7 @@ def __init__(
         self.activity_handles = []
         self.max_workers = max_workers
         self.max_concurrent_activities = max_concurrent_activities
+        self.max_cached_workflows = max_cached_workflows
         self.health_check_server_running = False
         self.healthy = False
         self.health_check_port = (
@@ -227,6 +229,7 @@ async def run(
             activities=activities,
             workflow_runner=UnsandboxedWorkflowRunner(),
             max_concurrent_activities=self.max_concurrent_activities,
+            max_cached_workflows=self.max_cached_workflows,
             build_id=str(uuid.uuid4()),
             debug_mode=debug_enabled,  # Disable deadlock detection in debug mode
             interceptors=self.interceptors,  # Pass interceptors to Worker