diff --git a/common/Definitions.h b/common/Definitions.h
index 45c43863..eca261df 100644
--- a/common/Definitions.h
+++ b/common/Definitions.h
@@ -49,9 +49,8 @@
 #define DEFAULT_G2_FACTOR   (-0.1145)
 
 // For FPGA
-/* This number is unique and is declared in ~snap/ActionTypes.md */
 #define ACTION_TYPE   0x52324158
-#define RELEASE_LEVEL 0x0035
+#define RELEASE_LEVEL 0x0036
 
 #define MODE_CONV                    0x0001L
 #define MODE_INTERNAL_PACKET_GEN     0x0002L
diff --git a/receiver/hls/host_writer.cpp b/receiver/hls/host_writer.cpp
index d7b03039..e73fca95 100644
--- a/receiver/hls/host_writer.cpp
+++ b/receiver/hls/host_writer.cpp
@@ -126,6 +126,15 @@ void host_writer(STREAM_512 &data_in,
         jf_bunchid[i]          = 0;
     }
 
+    ap_uint<32> req_handle;
+    ap_uint<64> req_host_offset;
+
+    while (data_in.empty()) {
+#pragma HLS PIPELINE II=4
+        if (!s_axis_work_request.empty())
+            read_request(s_axis_work_request, req_handle, req_host_offset);
+    }
+
     ap_uint<ADDR_STREAM_WIDTH> addr;
     addr_in >> addr;
 
@@ -150,9 +159,6 @@ void host_writer(STREAM_512 &data_in,
     packet_out.dest = 0;
     packet_out.id = 0;
     packet_out.user = 0;
-
-    ap_uint<32> req_handle;
-    ap_uint<64> req_host_offset;
     
     Loop_good_packet:
     while (!addr_last_flag(addr)) {
@@ -252,11 +258,4 @@ void host_writer(STREAM_512 &data_in,
     data_in >> packet_in;
 
     write_completion(m_axis_completion, HANDLE_END, 0, total_counter, 0, 0, 0, 0, 0, 0, data_collection_counter);
-
-    read_request(s_axis_work_request, req_handle, req_host_offset);
-
-    while (req_handle != UINT32_MAX) {
-#pragma HLS PIPELINE II=4
-        read_request(s_axis_work_request, req_handle, req_host_offset);
-    }
 }
diff --git a/receiver/host/AcquisitionDevice.cpp b/receiver/host/AcquisitionDevice.cpp
index e21c59ed..d4cda054 100644
--- a/receiver/host/AcquisitionDevice.cpp
+++ b/receiver/host/AcquisitionDevice.cpp
@@ -130,6 +130,8 @@ void AcquisitionDevice::StartAction(const DiffractionExperiment &experiment) {
     if (c.type != Completion::Type::Start)
         throw JFJochException(JFJochExceptionCategory::AcquisitionDeviceError, "Mismatch in completion queue");
 
+    StartSendingWorkRequests();
+
     start_time = std::chrono::system_clock::now();
 
     if (logger)
diff --git a/receiver/host/AcquisitionDevice.h b/receiver/host/AcquisitionDevice.h
index c8b7eb58..5a15e697 100644
--- a/receiver/host/AcquisitionDevice.h
+++ b/receiver/host/AcquisitionDevice.h
@@ -40,8 +40,6 @@ class AcquisitionDevice {
     AcquisitionOnlineCounters counters;
     AcquisitionOfflineCounters completion_vector;
 
-    void EndWorkRequestAndSignalQueues();
-
     virtual void HW_WriteActionRegister(const ActionConfig *job) = 0;
     virtual void HW_ReadActionRegister(ActionConfig *job) = 0;
     virtual void HW_StartAction() = 0;
@@ -56,6 +54,7 @@ class AcquisitionDevice {
     virtual uint64_t HW_GetMACAddress() const = 0;
     virtual void HW_EndAction() {}; // do clean-up after action is done
     virtual void CopyInternalPacketGenFrameToDeviceBuffer();
+    virtual void StartSendingWorkRequests() {};
 protected:
     ThreadSafeFIFO<Completion> work_completion_queue;
     ThreadSafeFIFO<WorkRequest> work_request_queue;
diff --git a/receiver/host/FPGAAcquisitionDevice.cpp b/receiver/host/FPGAAcquisitionDevice.cpp
index e526639b..6455b2d4 100644
--- a/receiver/host/FPGAAcquisitionDevice.cpp
+++ b/receiver/host/FPGAAcquisitionDevice.cpp
@@ -5,11 +5,12 @@
 
 void FPGAAcquisitionDevice::HW_StartAction() {
     FPGA_StartAction();
+    read_work_completion_future = std::async(std::launch::async, &FPGAAcquisitionDevice::ReadWorkCompletionThread, this);
+}
 
+void FPGAAcquisitionDevice::StartSendingWorkRequests() {
     stop_work_requests = false;
     send_work_request_future = std::async(std::launch::async, &FPGAAcquisitionDevice::SendWorkRequestThread, this);
-
-    read_work_completion_future = std::async(std::launch::async, &FPGAAcquisitionDevice::ReadWorkCompletionThread, this);
 }
 
 void FPGAAcquisitionDevice::HW_EndAction() {
@@ -18,9 +19,6 @@ void FPGAAcquisitionDevice::HW_EndAction() {
     stop_work_requests = true;
     send_work_request_future.get();
 
-    while (!HW_SendWorkRequest(UINT32_MAX))
-        std::this_thread::sleep_for(std::chrono::microseconds(10));
-
     FPGA_EndAction();
 }
 
diff --git a/receiver/host/FPGAAcquisitionDevice.h b/receiver/host/FPGAAcquisitionDevice.h
index 4df98d2c..831c032d 100644
--- a/receiver/host/FPGAAcquisitionDevice.h
+++ b/receiver/host/FPGAAcquisitionDevice.h
@@ -22,6 +22,7 @@ class FPGAAcquisitionDevice : public AcquisitionDevice {
 
     virtual bool HW_ReadMailbox(uint32_t values[16]) = 0;
     virtual bool HW_SendWorkRequest(uint32_t handle) = 0;
+    void StartSendingWorkRequests() override;
 protected:
     explicit FPGAAcquisitionDevice(uint16_t data_stream) : AcquisitionDevice(data_stream) {}
 };
diff --git a/receiver/host/HLSSimulatedDevice.cpp b/receiver/host/HLSSimulatedDevice.cpp
index ac23e2b2..b488f495 100644
--- a/receiver/host/HLSSimulatedDevice.cpp
+++ b/receiver/host/HLSSimulatedDevice.cpp
@@ -352,10 +352,7 @@ void HLSSimulatedDevice::HLSMainThread() {
     if (!datamover_in.GetDataStream().empty())
         throw std::runtime_error("Datamover queue is not empty");
 
-    if (!work_request_stream.empty())
-        throw std::runtime_error("Work request stream is not empty");
-
-    if (err_reg != 0)
+      if (err_reg != 0)
         throw std::runtime_error("Error reg for frame_statistics not zero, val=" + std::to_string(err_reg));
 
     while (!datamover_out.IsIdle())