diff --git a/agent/node_pool.py b/agent/node_pool.py
new file mode 100644
index 0000000..ef2e0cd
--- /dev/null
+++ b/agent/node_pool.py
@@ -0,0 +1,50 @@
+"""NodePool: shared stateless node instances across all sessions.
+
+Stateless nodes (InputNode, PANode, ExpertNode, etc.) hold no per-session
+state — only config (model, system prompt). They can safely serve multiple
+concurrent sessions. Session-specific HUD routing uses contextvars.
+
+Stateful nodes (SensorNode, MemorizerNode, UINode) hold conversational
+state and must be created per-session.
+"""
+
+import logging
+
+from .engine import load_graph, instantiate_nodes
+
+log = logging.getLogger("runtime")
+
+# Roles that hold per-session state — always created fresh per Runtime
+STATEFUL_ROLES = frozenset({"sensor", "memorizer", "ui"})
+
+
+async def _noop_hud(data: dict):
+    """Placeholder HUD — shared nodes use contextvars for session routing."""
+    pass
+
+
+class NodePool:
+    """Shared node instances for stateless LLM nodes.
+
+    Usage:
+        pool = NodePool("v4-eras")
+        # Shared nodes (one instance, all sessions):
+        input_node = pool.shared["input"]
+        # Stateful nodes must be created per-session (not in pool)
+    """
+
+    def __init__(self, graph_name: str = "v4-eras"):
+        self.graph = load_graph(graph_name)
+        self.graph_name = graph_name
+
+        # Instantiate all nodes with noop HUD (shared nodes use contextvars)
+        all_nodes = instantiate_nodes(self.graph, send_hud=_noop_hud)
+
+        # Split: shared (stateless) vs excluded (stateful)
+        self.shared = {
+            role: node for role, node in all_nodes.items()
+            if role not in STATEFUL_ROLES
+        }
+
+        log.info(f"[pool] created for graph '{graph_name}': "
+                 f"{len(self.shared)} shared, {len(STATEFUL_ROLES)} stateful")
diff --git a/agent/nodes/base.py b/agent/nodes/base.py
index c416330..396cb43 100644
--- a/agent/nodes/base.py
+++ b/agent/nodes/base.py
@@ -1,11 +1,16 @@
 """Base Node class with context management."""
 
+import contextvars
 import logging
 
 from ..llm import estimate_tokens, fit_context
 
 log = logging.getLogger("runtime")
 
+# Per-task HUD callback — set by FrameEngine/Runtime before calling shared nodes.
+# Isolates HUD events between concurrent sessions (asyncio.Task-scoped).
+_current_hud = contextvars.ContextVar('send_hud', default=None)
+
 
 class Node:
     name: str = "node"
@@ -18,10 +23,11 @@ class Node:
         self.context_fill_pct = 0
 
     async def hud(self, event: str, **data):
-        # Always include model on context events so frontend knows what model each node uses
+        # Use task-scoped HUD if set (shared node pool), else instance callback
+        hud_fn = _current_hud.get() or self.send_hud
         if event == "context" and self.model:
             data["model"] = self.model
-        await self.send_hud({"node": self.name, "event": event, **data})
+        await hud_fn({"node": self.name, "event": event, **data})
 
     def trim_context(self, messages: list[dict]) -> list[dict]:
         """Fit messages within this node's token budget."""