diff --git a/telethon/network/mtprotolayer.py b/telethon/network/mtprotolayer.py
index 754df8b5..8142a5e9 100644
--- a/telethon/network/mtprotolayer.py
+++ b/telethon/network/mtprotolayer.py
@@ -33,17 +33,16 @@ class MTProtoLayer:
         """
         self._connection.disconnect()
 
-    async def send(self, data_list):
+    async def send(self, state_list):
         """
-        A list of serialized RPC queries as bytes must be given to be sent.
+        The list of `RequestState` that will be sent. They will
+        be updated with their new message and container IDs.
+
         Nested lists imply an order is required for the messages in them.
         Message containers will be used if there is more than one item.
-
-        Returns ``(container_id, msg_ids)``.
         """
-        data, container_id, msg_ids = self._pack_data_list(data_list)
+        data = self._pack_state_list(state_list)
         await self._connection.send(self._state.encrypt_message_data(data))
-        return container_id, msg_ids
 
     async def recv(self):
         """
@@ -52,12 +51,14 @@ class MTProtoLayer:
         body = await self._connection.recv()
         return self._state.decrypt_message_data(body)
 
-    def _pack_data_list(self, data_list):
+    def _pack_state_list(self, state_list):
         """
-        A list of serialized RPC queries as bytes must be given to be packed.
-        Nested lists imply an order is required for the messages in them.
+        The list of `RequestState` that will be sent. They will
+        be updated with their new message and container IDs.
 
-        Returns ``(data, container_id, msg_ids)``.
+        Packs all their serialized data into a message (possibly
+        nested inside another message and message container) and
+        returns the serialized message data.
         """
         # TODO write_data_as_message raises on invalid messages, handle it
         # TODO This method could be an iterator yielding messages while small
@@ -72,33 +73,39 @@ class MTProtoLayer:
         # to store and serialize the data. However, to keep the context local
         # and relevant to the only place where such feature is actually used,
         # this is not done.
-        msg_ids = []
+        n = 0
         buffer = io.BytesIO()
-        for data in data_list:
-            if not isinstance(data, list):
-                msg_ids.append(self._state.write_data_as_message(buffer, data))
+        for state in state_list:
+            if not isinstance(state, list):
+                n += 1
+                state.msg_id = \
+                    self._state.write_data_as_message(buffer, state.data)
             else:
                 last_id = None
-                for d in data:
-                    last_id = self._state.write_data_as_message(
-                        buffer, d, after_id=last_id)
-                    msg_ids.append(last_id)
+                for s in state:
+                    n += 1
+                    last_id = s.msg_id = self._state.write_data_as_message(
+                        buffer, s.data, after_id=last_id)
 
-        if len(msg_ids) == 1:
-            container_id = None
-        else:
+        if n > 1:
             # Inlined code to pack several messages into a container
             #
             # TODO This part and encrypting data prepend a few bytes but
             # force a potentially large payload to be appended, which
             # may be expensive. Can we do better?
             data = struct.pack(
-                '<Ii', MessageContainer.CONSTRUCTOR_ID, len(msg_ids)
+                '<Ii', MessageContainer.CONSTRUCTOR_ID, n
             ) + buffer.getvalue()
             buffer = io.BytesIO()
             container_id = self._state.write_data_as_message(buffer, data)
+            for state in state_list:
+                if not isinstance(state, list):
+                    state.container_id = container_id
+                else:
+                    for s in state:
+                        s.container_id = container_id
 
-        return buffer.getvalue(), container_id, msg_ids
+        return buffer.getvalue()
 
     def __str__(self):
         return str(self._connection)
diff --git a/telethon/network/mtprotosender.py b/telethon/network/mtprotosender.py
index b5e25b5c..b51fa122 100644
--- a/telethon/network/mtprotosender.py
+++ b/telethon/network/mtprotosender.py
@@ -1,4 +1,5 @@
 import asyncio
+import collections
 import logging
 
 from .mtprotolayer import MTProtoLayer
@@ -15,16 +16,11 @@ from ..tl.types import (
     MsgNewDetailedInfo, NewSessionCreated, MsgDetailedInfo, MsgsStateReq,
     MsgsStateInfo, MsgsAllInfo, MsgResendReq, upload
 )
+from .requeststate import RequestState
 
 __log__ = logging.getLogger(__name__)
 
 
-# Place this object in the send queue when a reconnection is needed
-# so there is an item to read and we can early quit the loop, since
-# without this it will block until there's something in the queue.
-_reconnect_sentinel = object()
-
-
 class MTProtoSender:
     """
     MTProto Mobile Protocol sender
@@ -66,27 +62,21 @@ class MTProtoSender:
         self._send_loop_handle = None
         self._recv_loop_handle = None
 
-        # Sending something shouldn't block
-        self._send_queue = _ContainerQueue()
+        # Outgoing messages are put in a queue and sent in a batch.
+        # Note that here we're also storing their ``_RequestState``.
+        # Note that it may also store lists (implying order must be kept).
+        self._send_queue = asyncio.Queue()
 
-        # Telegram responds to messages out of order. Keep
-        # {id: Message} to set their Future result upon arrival.
-        self._pending_messages = {}
+        # Sent states are remembered until a response is received.
+        self._pending_state = {}
 
-        # Containers are accepted or rejected as a whole when any of
-        # its inner requests are acknowledged. For this purpose we
-        # all the sent containers here.
-        self._pending_containers = []
-
-        # We need to acknowledge every response from Telegram
+        # Responses must be acknowledged, and we can also batch these.
         self._pending_ack = set()
 
-        # Similar to pending_messages but only for the last ack.
-        # Ack can't be put in the messages because Telegram never
-        # responds to acknowledges (they're just that, acknowledges),
-        # so it would grow to infinite otherwise, but on bad salt it's
-        # necessary to resend them just like everything else.
-        self._last_ack = None
+        # Similar to pending_messages but only for the last acknowledges.
+        # These can't go in pending_messages because no acknowledge for them
+        # is received, but we may still need to resend their state on bad salts.
+        self._last_acks = collections.deque(maxlen=10)
 
         # Jump table from response ID to method that handles it
         self._handlers = {
@@ -136,6 +126,7 @@ class MTProtoSender:
 
         await self._disconnect()
 
+    # TODO Move this out of the "Public API" section
     async def _disconnect(self, error=None):
         __log__.info('Disconnecting from %s...', self._connection)
         self._user_connected = False
@@ -144,14 +135,14 @@ class MTProtoSender:
             self._connection.disconnect()
         finally:
             __log__.debug('Cancelling {} pending message(s)...'
-                          .format(len(self._pending_messages)))
-            for message in self._pending_messages.values():
-                if error and not message.future.done():
-                    message.future.set_exception(error)
+                          .format(len(self._pending_state)))
+            for state in self._pending_state.values():
+                if error and not state.future.done():
+                    state.future.set_exception(error)
                 else:
-                    message.future.cancel()
+                    state.future.cancel()
 
-            self._pending_messages.clear()
+            self._pending_state.clear()
             self._pending_ack.clear()
             self._last_ack = None
 
@@ -172,11 +163,9 @@ class MTProtoSender:
 
     def send(self, request, ordered=False):
         """
-        This method enqueues the given request to be sent.
-
-        The request will be wrapped inside a `TLMessage` until its
-        response arrives, and the `Future` response of the `TLMessage`
-        is immediately returned so that one can further ``await`` it:
+        This method enqueues the given request to be sent. Its send
+        state will be saved until a response arrives, and a ``Future``
+        that will be resolved when the response arrives will be returned:
 
         .. code-block:: python
 
@@ -198,23 +187,23 @@ class MTProtoSender:
         if not self._user_connected:
             raise ConnectionError('Cannot send requests while disconnected')
 
-        if utils.is_list_like(request):
-            result = []
-            after = None
-            for r in request:
-                message = self.state.create_message(
-                    r, loop=self._loop, after=after)
-
-                self._pending_messages[message.msg_id] = message
-                self._send_queue.put_nowait(message)
-                result.append(message.future)
-                after = ordered and message
-            return result
+        if not utils.is_list_like(request):
+            state = RequestState(request, self._loop)
+            self._send_queue.put_nowait(state)
+            return state.future
         else:
-            message = self.state.create_message(request, loop=self._loop)
-            self._pending_messages[message.msg_id] = message
-            self._send_queue.put_nowait(message)
-            return message.future
+            states = []
+            futures = []
+            for req in request:
+                state = RequestState(req, self._loop)
+                states.append(state)
+                futures.append(state.future)
+            if ordered:
+                self._send_queue.put(states)
+            else:
+                for state in states:
+                    self._send_queue.put(state)
+            return futures
 
     @property
     def disconnected(self):
@@ -290,7 +279,6 @@ class MTProtoSender:
         Cleanly disconnects and then reconnects.
         """
         self._reconnecting = True
-        self._send_queue.put_nowait(_reconnect_sentinel)
 
         __log__.debug('Awaiting for the send loop before reconnecting...')
         await self._send_loop_handle
@@ -307,8 +295,11 @@ class MTProtoSender:
         for retry in range(1, retries + 1):
             try:
                 await self._connect()
+                # TODO Keep this?
+                """
                 for m in self._pending_messages.values():
                     self._send_queue.put_nowait(m)
+                """
 
                 if self._auto_reconnect_callback:
                     self._loop.create_task(self._auto_reconnect_callback())
@@ -325,23 +316,6 @@ class MTProtoSender:
         if self._user_connected:
             self._loop.create_task(self._reconnect())
 
-    def _clean_containers(self, msg_ids):
-        """
-        Helper method to clean containers from the pending messages
-        once a wrapped msg_id of them has been acknowledged.
-
-        This is the only way we can resend TLMessage(MessageContainer)
-        on bad notifications and also mark them as received once any
-        of their inner TLMessage is acknowledged.
-        """
-        for i in reversed(range(len(self._pending_containers))):
-            message = self._pending_containers[i]
-            for msg in message.obj.messages:
-                if msg.msg_id in msg_ids:
-                    del self._pending_containers[i]
-                    del self._pending_messages[message.msg_id]
-                    break
-
     # Loops
 
     async def _send_loop(self):
@@ -353,67 +327,31 @@ class MTProtoSender:
         """
         while self._user_connected and not self._reconnecting:
             if self._pending_ack:
-                self._last_ack = self.state.create_message(
-                    MsgsAck(list(self._pending_ack)), loop=self._loop
-                )
-                self._send_queue.put_nowait(self._last_ack)
+                ack = RequestState(MsgsAck(list(self._pending_ack)), self._loop)
+                self._send_queue.put_nowait(ack)
+                self._last_acks.append(ack)
                 self._pending_ack.clear()
 
-            messages = await self._send_queue.get()
-            if messages == _reconnect_sentinel:
-                if self._reconnecting:
-                    break
+            state_list = []
+
+            # TODO wait for the list to have one or for a disconnect to happen
+            # and pop while that's the case
+            state = await self._send_queue.get()
+            state_list.append(state)
+
+            while not self._send_queue.empty():
+                state_list.append(self._send_queue.get_nowait())
+
+            # TODO Debug logs to notify which messages are being sent
+            # TODO Try sending them while no future was cancelled?
+            # TODO Handle timeout, cancelled, arbitrary errors
+            await self._connection.send(state_list)
+            for state in state_list:
+                if not isinstance(state, list):
+                    self._pending_state[state.msg_id] = state
                 else:
-                    continue
-
-            if isinstance(messages, list):
-                message = self.state.create_message(
-                    MessageContainer(messages), loop=self._loop)
-
-                self._pending_messages[message.msg_id] = message
-                self._pending_containers.append(message)
-            else:
-                message = messages
-                messages = [message]
-
-            __log__.debug(
-                'Packing %d outgoing message(s) %s...', len(messages),
-                ', '.join(x.obj.__class__.__name__ for x in messages)
-            )
-            body = self.state.pack_message(message)
-
-            while not any(m.future.cancelled() for m in messages):
-                try:
-                    __log__.debug('Sending {} bytes...'.format(len(body)))
-                    await self._connection.send(body)
-                    break
-                except asyncio.TimeoutError:
-                    continue
-                except asyncio.CancelledError:
-                    return
-                except Exception as e:
-                    if isinstance(e, ConnectionError):
-                        __log__.info('Connection reset while sending %s', e)
-                    elif isinstance(e, OSError):
-                        __log__.warning('OSError while sending %s', e)
-                    else:
-                        __log__.exception('Unhandled exception while receiving')
-                        await asyncio.sleep(1, loop=self._loop)
-
-                    self._start_reconnect()
-                    break
-            else:
-                # Remove the cancelled messages from pending
-                __log__.info('Some futures were cancelled, aborted send')
-                self._clean_containers([m.msg_id for m in messages])
-                for m in messages:
-                    if m.future.cancelled():
-                        self._pending_messages.pop(m.msg_id, None)
-                    else:
-                        self._send_queue.put_nowait(m)
-
-            __log__.debug('Outgoing messages {} sent!'
-                          .format(', '.join(str(m.msg_id) for m in messages)))
+                    for s in state:
+                        self._pending_state[s.msg_id] = s
 
     async def _recv_loop(self):
         """
@@ -423,68 +361,11 @@ class MTProtoSender:
         Besides `connect`, only this method ever receives data.
         """
         while self._user_connected and not self._reconnecting:
-            try:
-                __log__.debug('Receiving items from the network...')
-                body = await self._connection.recv()
-            except asyncio.TimeoutError:
-                continue
-            except asyncio.CancelledError:
-                return
-            except Exception as e:
-                if isinstance(e, ConnectionError):
-                    __log__.info('Connection reset while receiving %s', e)
-                elif isinstance(e, OSError):
-                    __log__.warning('OSError while receiving %s', e)
-                else:
-                    __log__.exception('Unhandled exception while receiving')
-                    await asyncio.sleep(1, loop=self._loop)
-
-                self._start_reconnect()
-                break
-
-            # TODO Check salt, session_id and sequence_number
-            __log__.debug('Decoding packet of %d bytes...', len(body))
-            try:
-                message = self.state.unpack_message(body)
-            except (BrokenAuthKeyError, BufferError) as e:
-                # The authorization key may be broken if a message was
-                # sent malformed, or if the authkey truly is corrupted.
-                #
-                # There may be a buffer error if Telegram's response was too
-                # short and hence not understood. Reset the authorization key
-                # and try again in either case.
-                #
-                # TODO Is it possible to detect malformed messages vs
-                # an actually broken authkey?
-                __log__.warning('Broken authorization key?: {}'.format(e))
-                self.state.auth_key = None
-                self._start_reconnect()
-                break
-            except SecurityError as e:
-                # A step while decoding had the incorrect data. This message
-                # should not be considered safe and it should be ignored.
-                __log__.warning('Security error while unpacking a '
-                                'received message: {}'.format(e))
-                continue
-            except TypeNotFoundError as e:
-                # The payload inside the message was not a known TLObject.
-                __log__.info('Server replied with an unknown type {:08x}: {!r}'
-                             .format(e.invalid_constructor_id, e.remaining))
-                continue
-            except asyncio.CancelledError:
-                return
-            except Exception as e:
-                __log__.exception('Unhandled exception while unpacking %s',e)
-                await asyncio.sleep(1, loop=self._loop)
-            else:
-                try:
-                    await self._process_message(message)
-                except asyncio.CancelledError:
-                    return
-                except Exception as e:
-                    __log__.exception('Unhandled exception while '
-                                      'processing %s', message)
-                    await asyncio.sleep(1, loop=self._loop)
+            # TODO Handle timeout, cancelled, arbitrary, broken auth, buffer,
+            # security and type not found.
+            __log__.debug('Receiving items from the network...')
+            message = await self._connection.recv()
+            await self._process_message(message)
 
     # Response Handlers
 
@@ -508,11 +389,11 @@ class MTProtoSender:
         This is where the future results for sent requests are set.
         """
         rpc_result = message.obj
-        message = self._pending_messages.pop(rpc_result.req_msg_id, None)
+        state = self._pending_state.pop(rpc_result.req_msg_id, None)
         __log__.debug('Handling RPC result for message %d',
                       rpc_result.req_msg_id)
 
-        if not message:
+        if not state:
             # TODO We should not get responses to things we never sent
             # However receiving a File() with empty bytes is "common".
             # See #658, #759 and #958. They seem to happen in a container
@@ -528,22 +409,21 @@ class MTProtoSender:
 
         if rpc_result.error:
             error = rpc_message_to_error(rpc_result.error)
-            self._send_queue.put_nowait(self.state.create_message(
-                MsgsAck([message.msg_id]), loop=self._loop
-            ))
+            self._send_queue.put_nowait(
+                RequestState(MsgsAck([state.msg_id]), loop=self._loop))
 
-            if not message.future.cancelled():
-                message.future.set_exception(error)
+            if not state.future.cancelled():
+                state.future.set_exception(error)
         else:
             # TODO Would be nice to avoid accessing a per-obj read_result
             # Instead have a variable that indicated how the result should
             # be read (an enum) and dispatch to read the result, mostly
             # always it's just a normal TLObject.
             with BinaryReader(rpc_result.body) as reader:
-                result = message.obj.read_result(reader)
+                result = state.request.read_result(reader)
 
-            if not message.future.cancelled():
-                message.future.set_result(result)
+            if not state.future.cancelled():
+                state.future.set_result(result)
 
     async def _handle_container(self, message):
         """
@@ -581,9 +461,9 @@ class MTProtoSender:
         """
         pong = message.obj
         __log__.debug('Handling pong for message %d', pong.msg_id)
-        message = self._pending_messages.pop(pong.msg_id, None)
-        if message:
-            message.future.set_result(pong)
+        state = self._pending_state.pop(pong.msg_id, None)
+        if state:
+            state.future.set_result(pong)
 
     async def _handle_bad_server_salt(self, message):
         """
@@ -595,16 +475,16 @@ class MTProtoSender:
         """
         bad_salt = message.obj
         __log__.debug('Handling bad salt for message %d', bad_salt.bad_msg_id)
-        self.state.salt = bad_salt.new_server_salt
-        if self._last_ack and bad_salt.bad_msg_id == self._last_ack.msg_id:
-            self._send_queue.put_nowait(self._last_ack)
-            return
-
+        self._connection._state.salt = bad_salt.new_server_salt
         try:
             self._send_queue.put_nowait(
-                self._pending_messages[bad_salt.bad_msg_id])
+                self._pending_state.pop(bad_salt.bad_msg_id))
         except KeyError:
-            # May be MsgsAck, those are not saved in pending messages
+            for ack in self._pending_ack:
+                if ack.msg_id == bad_salt.bad_msg_id:
+                    self._send_queue.put_nowait(ack)
+                    return
+
             __log__.info('Message %d not resent due to bad salt',
                          bad_salt.bad_msg_id)
 
@@ -617,41 +497,33 @@ class MTProtoSender:
             error_code:int = BadMsgNotification;
         """
         bad_msg = message.obj
-        msg = self._pending_messages.get(bad_msg.bad_msg_id)
+        # TODO Pending state may need to pop by container ID
+        state = self._pending_state.pop(bad_msg.bad_msg_id, None)
 
         __log__.debug('Handling bad msg %s', bad_msg)
         if bad_msg.error_code in (16, 17):
             # Sent msg_id too low or too high (respectively).
             # Use the current msg_id to determine the right time offset.
-            to = self.state.update_time_offset(correct_msg_id=message.msg_id)
+            to = self._connection._state.update_time_offset(
+                correct_msg_id=message.msg_id)
             __log__.info('System clock is wrong, set time offset to %ds', to)
-
-            # Correct the msg_id *of the message to resend*, not all.
-            #
-            # If we correct them all, new "bad message" would not find
-            # the old invalid IDs, causing all awaits to never finish.
-            if msg:
-                del self._pending_messages[msg.msg_id]
-                self.state.update_message_id(msg)
-                self._pending_messages[msg.msg_id] = msg
-
         elif bad_msg.error_code == 32:
             # msg_seqno too low, so just pump it up by some "large" amount
             # TODO A better fix would be to start with a new fresh session ID
-            self.state._sequence += 64
+            self._connection._state._sequence += 64
         elif bad_msg.error_code == 33:
             # msg_seqno too high never seems to happen but just in case
-            self.state._sequence -= 16
+            self._connection._state._sequence -= 16
         else:
-            if msg:
-                del self._pending_messages[msg.msg_id]
-                msg.future.set_exception(BadMessageError(bad_msg.error_code))
+            if state:
+                state.future.set_exception(BadMessageError(bad_msg.error_code))
             return
 
         # Messages are to be re-sent once we've corrected the issue
-        if msg:
-            self._send_queue.put_nowait(msg)
+        if state:
+            self._send_queue.put_nowait(state)
         else:
+            # TODO Generic method that may return from the acks too
             # May be MsgsAck, those are not saved in pending messages
             __log__.info('Message %d not resent due to bad msg',
                          bad_msg.bad_msg_id)
@@ -689,7 +561,7 @@ class MTProtoSender:
         """
         # TODO https://goo.gl/LMyN7A
         __log__.debug('Handling new session created')
-        self.state.salt = message.obj.server_salt
+        self._connection._state.salt = message.obj.server_salt
 
     async def _handle_ack(self, message):
         """
@@ -708,14 +580,11 @@ class MTProtoSender:
         """
         ack = message.obj
         __log__.debug('Handling acknowledge for %s', str(ack.msg_ids))
-        if self._pending_containers:
-            self._clean_containers(ack.msg_ids)
-
         for msg_id in ack.msg_ids:
-            msg = self._pending_messages.get(msg_id, None)
-            if msg and isinstance(msg.obj, LogOutRequest):
-                del self._pending_messages[msg_id]
-                msg.future.set_result(True)
+            state = self._pending_state.get(msg_id)
+            if state and isinstance(state.request, LogOutRequest):
+                del self._pending_state[msg_id]
+                state.future.set_result(True)
 
     async def _handle_future_salts(self, message):
         """
@@ -728,52 +597,20 @@ class MTProtoSender:
         # TODO save these salts and automatically adjust to the
         # correct one whenever the salt in use expires.
         __log__.debug('Handling future salts for message %d', message.msg_id)
-        msg = self._pending_messages.pop(message.msg_id, None)
-        if msg:
-            msg.future.set_result(message.obj)
+        state = self._pending_state.pop(message.msg_id, None)
+        if state:
+            state.future.set_result(message.obj)
 
     async def _handle_state_forgotten(self, message):
         """
         Handles both :tl:`MsgsStateReq` and :tl:`MsgResendReq` by
         enqueuing a :tl:`MsgsStateInfo` to be sent at a later point.
         """
-        self.send(MsgsStateInfo(req_msg_id=message.msg_id,
-                                info=chr(1) * len(message.obj.msg_ids)))
+        self._send_queue.put_nowait(RequestState(MsgsStateInfo(
+            req_msg_id=message.msg_id, info=chr(1) * len(message.obj.msg_ids)),
+            loop=self._loop))
 
     async def _handle_msg_all(self, message):
         """
         Handles :tl:`MsgsAllInfo` by doing nothing (yet).
         """
-
-
-class _ContainerQueue(asyncio.Queue):
-    """
-    An asyncio queue that's aware of `MessageContainer` instances.
-
-    The `get` method returns either a single `TLMessage` or a list
-    of them that should be turned into a new `MessageContainer`.
-
-    Instances of this class can be replaced with the simpler
-    ``asyncio.Queue`` when needed for testing purposes, and
-    a list won't be returned in said case.
-    """
-    async def get(self):
-        result = await super().get()
-        if self.empty() or result == _reconnect_sentinel or\
-                isinstance(result.obj, MessageContainer):
-            return result
-
-        size = result.size()
-        result = [result]
-        while not self.empty():
-            item = self.get_nowait()
-            if (item == _reconnect_sentinel or
-                isinstance(item.obj, MessageContainer)
-                    or size + item.size() > MessageContainer.MAXIMUM_SIZE):
-                self.put_nowait(item)
-                break
-            else:
-                size += item.size()
-                result.append(item)
-
-        return result
diff --git a/telethon/network/requeststate.py b/telethon/network/requeststate.py
new file mode 100644
index 00000000..018832af
--- /dev/null
+++ b/telethon/network/requeststate.py
@@ -0,0 +1,18 @@
+import asyncio
+
+
+class RequestState:
+    """
+    This request state holds several information relevant to sent messages,
+    in particular the message ID assigned to the request, the container ID
+    it belongs to, the request itself, the request as bytes, and the future
+    result that will eventually be resolved.
+    """
+    __slots__ = ('container_id', 'msg_id', 'request', 'data', 'future')
+
+    def __init__(self, request, loop):
+        self.container_id = None
+        self.msg_id = None
+        self.request = request
+        self.data = bytes(request)
+        self.future = asyncio.Future(loop=loop)