diff --git a/base/task.jl b/base/task.jl
index db586d59d02ca..f9adfb9d418fe 100644
--- a/base/task.jl
+++ b/base/task.jl
@@ -406,6 +406,12 @@ isolating the asynchronous code from changes to the variable's value in the curr
     Interpolating values via `\$` is available as of Julia 1.4.
 """
 macro async(expr)
+    do_async_macro(expr)
+end
+
+# generate the code for @async, possibly wrapping the task in something before
+# pushing it to the wait queue.
+function do_async_macro(expr; wrap=identity)
     letargs = Base._lift_one_interp!(expr)
 
     thunk = esc(:(()->($expr)))
@@ -414,7 +420,7 @@ macro async(expr)
         let $(letargs...)
             local task = Task($thunk)
             if $(Expr(:islocal, var))
-                put!($var, task)
+                put!($var, $(wrap(:task)))
             end
             schedule(task)
             task
@@ -422,6 +428,35 @@ macro async(expr)
     end
 end
 
+# task wrapper that doesn't create exceptions wrapped in TaskFailedException
+struct UnwrapTaskFailedException
+    task::Task
+end
+
+# common code for wait&fetch for UnwrapTaskFailedException
+function unwrap_task_failed(f::Function, t::UnwrapTaskFailedException)
+    try
+        f(t.task)
+    catch ex
+        if ex isa TaskFailedException
+            throw(ex.task.exception)
+        else
+            rethrow()
+        end
+    end
+end
+
+# the unwrapping for above task wrapper (gets triggered in sync_end())
+wait(t::UnwrapTaskFailedException) = unwrap_task_failed(wait, t)
+
+# same for fetching the tasks, for convenience
+fetch(t::UnwrapTaskFailedException) = unwrap_task_failed(fetch, t)
+
+# macro for running async code that doesn't throw wrapped exceptions
+macro async_unwrap(expr)
+    do_async_macro(expr, wrap=task->:(Base.UnwrapTaskFailedException($task)))
+end
+
 # Capture interpolated variables in $() and move them to let-block
 function _lift_one_interp!(e)
     letargs = Any[]  # store the new gensymed arguments
diff --git a/stdlib/Distributed/src/Distributed.jl b/stdlib/Distributed/src/Distributed.jl
index 50108f05eed26..4c7427e265afa 100644
--- a/stdlib/Distributed/src/Distributed.jl
+++ b/stdlib/Distributed/src/Distributed.jl
@@ -10,7 +10,7 @@ import Base: getindex, wait, put!, take!, fetch, isready, push!, length,
              hash, ==, kill, close, isopen, showerror
 
 # imports for use
-using Base: Process, Semaphore, JLOptions, buffer_writes, @sync_add,
+using Base: Process, Semaphore, JLOptions, buffer_writes, @async_unwrap,
             VERSION_STRING, binding_module, atexit, julia_exename,
             julia_cmd, AsyncGenerator, acquire, release, invokelatest,
             shell_escape_posixly, shell_escape_wincmd, escape_microsoft_c_args,
@@ -75,7 +75,7 @@ function _require_callback(mod::Base.PkgId)
         # broadcast top-level (e.g. from Main) import/using from node 1 (only)
         @sync for p in procs()
             p == 1 && continue
-            @sync_add remotecall(p) do
+            @async_unwrap remotecall_wait(p) do
                 Base.require(mod)
                 nothing
             end
diff --git a/stdlib/Distributed/src/clusterserialize.jl b/stdlib/Distributed/src/clusterserialize.jl
index e37987c5bf875..28025ae867c78 100644
--- a/stdlib/Distributed/src/clusterserialize.jl
+++ b/stdlib/Distributed/src/clusterserialize.jl
@@ -243,7 +243,7 @@ An exception is raised if a global constant is requested to be cleared.
 """
 function clear!(syms, pids=workers(); mod=Main)
     @sync for p in pids
-        @sync_add remotecall(clear_impl!, p, syms, mod)
+        @async_unwrap remotecall_wait(clear_impl!, p, syms, mod)
     end
 end
 clear!(sym::Symbol, pid::Int; mod=Main) = clear!([sym], [pid]; mod=mod)
diff --git a/stdlib/Distributed/src/macros.jl b/stdlib/Distributed/src/macros.jl
index b53890017d4de..f14aa7f42e3c1 100644
--- a/stdlib/Distributed/src/macros.jl
+++ b/stdlib/Distributed/src/macros.jl
@@ -226,10 +226,10 @@ function remotecall_eval(m::Module, procs, ex)
             if pid == myid()
                 run_locally += 1
             else
-                @sync_add remotecall(Core.eval, pid, m, ex)
+                @async_unwrap remotecall_wait(Core.eval, pid, m, ex)
             end
         end
-        yield() # ensure that the remotecall_fetch have had a chance to start
+        yield() # ensure that the remotecalls have had a chance to start
 
         # execute locally last as we do not want local execution to block serialization
         # of the request to remote nodes.