diff --git a/crates/discos-client/Cargo.toml b/crates/discos-client/Cargo.toml
index e4a2340..95a37c6 100644
--- a/crates/discos-client/Cargo.toml
+++ b/crates/discos-client/Cargo.toml
@@ -28,3 +28,7 @@ tempfile = "3"
 tokio-stream = "0.1"
 wasmparser = "0.220"
 wat = "1"
+
+[[test]]
+name = "agent_resilience"
+path = "../../tests/agent_resilience.rs"
diff --git a/tests/agent_resilience.rs b/tests/agent_resilience.rs
new file mode 100644
index 0000000..3d8000a
--- /dev/null
+++ b/tests/agent_resilience.rs
@@ -0,0 +1,290 @@
+// Copyright 2026 Joseph Verdicchio
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::sync::Arc;
+
+use tokio::sync::{mpsc, Mutex, Semaphore};
+use tokio::time::{sleep, Duration};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum KernelStatus {
+    Healthy,
+    Frozen,
+}
+
+#[derive(Debug, Clone)]
+struct FakeKernel {
+    state: Arc<Mutex<KernelState>>,
+}
+
+#[derive(Debug)]
+struct KernelState {
+    status: KernelStatus,
+    processed: usize,
+    max_inflight: usize,
+    inflight: usize,
+}
+
+impl FakeKernel {
+    fn new() -> Self {
+        Self {
+            state: Arc::new(Mutex::new(KernelState {
+                status: KernelStatus::Healthy,
+                processed: 0,
+                max_inflight: 0,
+                inflight: 0,
+            })),
+        }
+    }
+
+    async fn set_status(&self, status: KernelStatus) {
+        self.state.lock().await.status = status;
+    }
+
+    async fn submit_claim_capsule(&self, _capsule_id: u64) -> Result<(), ()> {
+        {
+            let mut state = self.state.lock().await;
+            if state.status == KernelStatus::Frozen {
+                return Err(());
+            }
+            state.inflight += 1;
+            state.max_inflight = state.max_inflight.max(state.inflight);
+        }
+
+        sleep(Duration::from_micros(50)).await;
+
+        let mut state = self.state.lock().await;
+        state.inflight -= 1;
+        state.processed += 1;
+        Ok(())
+    }
+
+    async fn snapshot(&self) -> KernelStateSnapshot {
+        let state = self.state.lock().await;
+        KernelStateSnapshot {
+            processed: state.processed,
+            max_inflight: state.max_inflight,
+        }
+    }
+}
+
+#[derive(Debug)]
+struct KernelStateSnapshot {
+    processed: usize,
+    max_inflight: usize,
+}
+
+#[derive(Debug)]
+struct AgentRuntimeState {
+    fail_safe: bool,
+    session_generation: u64,
+    frozen_session_generation: Option<u64>,
+    last_frozen_claim: Option<u64>,
+    max_buffered: usize,
+}
+
+#[derive(Debug, Clone)]
+struct AgentRuntime {
+    sender: mpsc::Sender<u64>,
+    state: Arc<Mutex<AgentRuntimeState>>,
+}
+
+impl AgentRuntime {
+    fn new(buffer: usize, worker_concurrency: usize, kernel: FakeKernel) -> Self {
+        let (sender, mut receiver) = mpsc::channel::<u64>(buffer);
+        let state = Arc::new(Mutex::new(AgentRuntimeState {
+            fail_safe: false,
+            session_generation: 1,
+            frozen_session_generation: None,
+            last_frozen_claim: None,
+            max_buffered: 0,
+        }));
+
+        let state_for_worker = Arc::clone(&state);
+        tokio::spawn(async move {
+            let permits = Arc::new(Semaphore::new(worker_concurrency));
+            let mut joins = Vec::new();
+
+            while let Some(capsule_id) = receiver.recv().await {
+                let permit = Arc::clone(&permits)
+                    .acquire_owned()
+                    .await
+                    .expect("worker semaphore should remain available");
+                let kernel_clone = kernel.clone();
+                let state_clone = Arc::clone(&state_for_worker);
+                joins.push(tokio::spawn(async move {
+                    let submit_result = kernel_clone.submit_claim_capsule(capsule_id).await;
+                    if submit_result.is_err() {
+                        let mut state = state_clone.lock().await;
+                        state.fail_safe = true;
+                        state.frozen_session_generation = Some(state.session_generation);
+                        state.last_frozen_claim = Some(capsule_id);
+                    }
+                    drop(permit);
+                }));
+            }
+
+            for handle in joins {
+                handle.await.expect("worker task must not panic");
+            }
+        });
+
+        Self { sender, state }
+    }
+
+    async fn enqueue_claim_capsule(&self, claim_id: u64) {
+        self.sender
+            .send(claim_id)
+            .await
+            .expect("queue should stay alive while runtime exists");
+
+        let buffered = self.sender.max_capacity() - self.sender.capacity();
+        let mut state = self.state.lock().await;
+        state.max_buffered = state.max_buffered.max(buffered);
+    }
+
+    async fn fail_safe(&self) -> bool {
+        self.state.lock().await.fail_safe
+    }
+
+    async fn queue_peak(&self) -> usize {
+        self.state.lock().await.max_buffered
+    }
+
+    async fn reauthenticate_after_ledger_reset(&self) {
+        let mut state = self.state.lock().await;
+        state.session_generation += 1;
+        state.fail_safe = false;
+        state.frozen_session_generation = None;
+        state.last_frozen_claim = None;
+    }
+
+    async fn snapshot(&self) -> AgentRuntimeStateSnapshot {
+        let state = self.state.lock().await;
+        AgentRuntimeStateSnapshot {
+            fail_safe: state.fail_safe,
+            session_generation: state.session_generation,
+            frozen_session_generation: state.frozen_session_generation,
+            last_frozen_claim: state.last_frozen_claim,
+        }
+    }
+}
+
+#[derive(Debug)]
+struct AgentRuntimeStateSnapshot {
+    fail_safe: bool,
+    session_generation: u64,
+    frozen_session_generation: Option<u64>,
+    last_frozen_claim: Option<u64>,
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn kernel_cutoff_frozen_status_enters_fail_safe() {
+    let kernel = FakeKernel::new();
+    kernel.set_status(KernelStatus::Frozen).await;
+    let runtime = AgentRuntime::new(8, 2, kernel);
+
+    runtime.enqueue_claim_capsule(42).await;
+    sleep(Duration::from_millis(20)).await;
+
+    let state = runtime.snapshot().await;
+    assert!(state.fail_safe, "runtime should enter fail-safe mode");
+    assert_eq!(state.frozen_session_generation, Some(1));
+    assert_eq!(state.last_frozen_claim, Some(42));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn heavy_probing_claim_capsules_applies_backpressure_without_drops() {
+    let kernel = FakeKernel::new();
+    let runtime = AgentRuntime::new(64, 32, kernel.clone());
+
+    let mut senders = Vec::with_capacity(5_000);
+    for claim_id in 0_u64..5_000 {
+        let runtime_clone = runtime.clone();
+        senders.push(tokio::spawn(async move {
+            runtime_clone.enqueue_claim_capsule(claim_id).await;
+        }));
+    }
+
+    for sender in senders {
+        sender.await.expect("sender task should not panic");
+    }
+
+    let queue_peak = runtime.queue_peak().await;
+    drop(runtime);
+    sleep(Duration::from_millis(250)).await;
+
+    let kernel_state = kernel.snapshot().await;
+    assert_eq!(
+        kernel_state.processed, 5_000,
+        "no claim capsule should be dropped"
+    );
+    assert!(
+        kernel_state.max_inflight <= 32,
+        "gRPC client concurrency must remain bounded"
+    );
+
+    assert!(
+        queue_peak <= 64,
+        "queue occupancy should be capped by channel capacity"
+    );
+    assert!(
+        queue_peak > 0,
+        "heavy probing should exercise queue backpressure"
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn reauthentication_clears_frozen_session_state_after_ledger_reset() {
+    let kernel = FakeKernel::new();
+    kernel.set_status(KernelStatus::Frozen).await;
+    let runtime = AgentRuntime::new(4, 1, kernel.clone());
+
+    runtime.enqueue_claim_capsule(7).await;
+    sleep(Duration::from_millis(20)).await;
+    assert!(
+        runtime.fail_safe().await,
+        "runtime should be frozen before reset"
+    );
+
+    kernel.set_status(KernelStatus::Healthy).await;
+    runtime.reauthenticate_after_ledger_reset().await;
+
+    let state = runtime.snapshot().await;
+    assert!(
+        !state.fail_safe,
+        "fail-safe should be cleared after re-authentication"
+    );
+    assert_eq!(
+        state.session_generation, 2,
+        "session generation should rotate"
+    );
+    assert_eq!(
+        state.frozen_session_generation, None,
+        "frozen session marker should be dropped"
+    );
+    assert_eq!(
+        state.last_frozen_claim, None,
+        "frozen claim memory should be wiped"
+    );
+
+    runtime.enqueue_claim_capsule(8).await;
+    sleep(Duration::from_millis(20)).await;
+
+    let kernel_state = kernel.snapshot().await;
+    assert_eq!(
+        kernel_state.processed, 1,
+        "post-reset session must operate independently"
+    );
+}