diff --git a/lib/runtime/src/discovery/kube.rs b/lib/runtime/src/discovery/kube.rs
index 5fd65c9f30..c39439212b 100644
--- a/lib/runtime/src/discovery/kube.rs
+++ b/lib/runtime/src/discovery/kube.rs
@@ -19,14 +19,30 @@ use async_trait::async_trait;
 use kube::Client as KubeClient;
 use std::collections::HashSet;
 use std::sync::Arc;
-use tokio::sync::RwLock;
+use tokio::sync::{OnceCell, RwLock};
+use tokio::task::JoinHandle;
 
 /// Kubernetes-based discovery client
 #[derive(Clone)]
 pub struct KubeDiscoveryClient {
     instance_id: u64,
     metadata: Arc<RwLock<DiscoveryMetadata>>,
+    kube_client: KubeClient,
+    pod_info: PodInfo,
+    cancel_token: CancellationToken,
+    daemon_state: Arc<OnceCell<Arc<DaemonHandles>>>,
+}
+
+struct DaemonHandles {
     metadata_watch: tokio::sync::watch::Receiver<Arc<MetadataSnapshot>>,
+    #[allow(dead_code)]
+    daemon_handle: JoinHandle<()>,
+}
+
+impl DaemonHandles {
+    fn receiver(&self) -> tokio::sync::watch::Receiver<Arc<MetadataSnapshot>> {
+        self.metadata_watch.clone()
+    }
 }
 
 impl KubeDiscoveryClient {
@@ -53,26 +69,51 @@ impl KubeDiscoveryClient {
             .await
             .map_err(|e| anyhow::anyhow!("Failed to create Kubernetes client: {}", e))?;
 
-        // Create watch channel with initial empty snapshot
-        let (watch_tx, watch_rx) = tokio::sync::watch::channel(Arc::new(MetadataSnapshot::empty()));
-
-        // Create and spawn daemon
-        let daemon = DiscoveryDaemon::new(kube_client, pod_info, cancel_token)?;
-
-        tokio::spawn(async move {
-            if let Err(e) = daemon.run(watch_tx).await {
-                tracing::error!("Discovery daemon failed: {}", e);
-            }
-        });
-
-        tracing::info!("Discovery daemon started");
-
         Ok(Self {
             instance_id,
             metadata,
-            metadata_watch: watch_rx,
+            kube_client,
+            pod_info,
+            cancel_token,
+            daemon_state: Arc::new(OnceCell::new()),
         })
     }
+
+    async fn metadata_watch(&self) -> Result<tokio::sync::watch::Receiver<Arc<MetadataSnapshot>>> {
+        let handles = self
+            .daemon_state
+            .get_or_try_init(|| {
+                let instance_id = self.instance_id;
+                let kube_client = self.kube_client.clone();
+                let pod_info = self.pod_info.clone();
+                let cancel_token = self.cancel_token.clone();
+
+                async move {
+                    let (watch_tx, watch_rx) =
+                        tokio::sync::watch::channel(Arc::new(MetadataSnapshot::empty()));
+                    let daemon = DiscoveryDaemon::new(kube_client, pod_info, cancel_token)?;
+
+                    let daemon_handle = tokio::spawn(async move {
+                        if let Err(e) = daemon.run(watch_tx).await {
+                            tracing::error!("Discovery daemon failed: {}", e);
+                        }
+                    });
+
+                    tracing::info!(
+                        "Discovery daemon started lazily for instance_id={:x}",
+                        instance_id
+                    );
+
+                    Ok::<Arc<DaemonHandles>, anyhow::Error>(Arc::new(DaemonHandles {
+                        metadata_watch: watch_rx,
+                        daemon_handle,
+                    }))
+                }
+            })
+            .await?;
+
+        Ok(handles.receiver())
+    }
 }
 
 #[async_trait]
@@ -144,8 +185,31 @@ impl Discovery for KubeDiscoveryClient {
     async fn list(&self, query: DiscoveryQuery) -> Result<Vec<DiscoveryInstance>> {
         tracing::debug!("KubeDiscoveryClient::list called with query={:?}", query);
 
-        // Get current snapshot (may be empty if daemon hasn't fetched yet)
-        let snapshot = self.metadata_watch.borrow().clone();
+        // Ensure the daemon is running before accessing the snapshot
+        let mut metadata_watch = self.metadata_watch().await?;
+
+        // Check if we need to wait for initial snapshot.
+        // With lazy daemon start, the first call to list() triggers the daemon which needs time
+        // to poll other pods and aggregate their metadata before we have meaningful results.
+        let needs_wait = {
+            let snapshot = metadata_watch.borrow();
+            snapshot.sequence == 0 && snapshot.instances.is_empty()
+        };
+
+        // Wait for daemon to fetch at least one snapshot if this is the initial empty state
+        if needs_wait {
+            tracing::debug!("Waiting for initial discovery snapshot...");
+            // Wait for first update with a timeout
+            tokio::time::timeout(std::time::Duration::from_secs(10), metadata_watch.changed())
+                .await
+                .map_err(|_| anyhow::anyhow!("Timeout waiting for initial discovery snapshot"))?
+                .map_err(|_| {
+                    anyhow::anyhow!("Discovery daemon stopped before providing snapshot")
+                })?;
+        }
+
+        // Get current snapshot
+        let snapshot = metadata_watch.borrow().clone();
 
         tracing::debug!(
             "List using snapshot seq={} with {} instances",
@@ -177,8 +241,8 @@ impl Discovery for KubeDiscoveryClient {
             query
         );
 
-        // Clone the watch receiver
-        let mut watch_rx = self.metadata_watch.clone();
+        // Clone the watch receiver (starts daemon lazily on demand)
+        let mut watch_rx = self.metadata_watch().await?;
 
         // Create output stream
         let (event_tx, event_rx) = mpsc::unbounded_channel();