diff --git a/free/src/main/scala/cats/free/FreeT.scala b/free/src/main/scala/cats/free/FreeT.scala
index ffd06aa1c8..caa8b9e69d 100644
--- a/free/src/main/scala/cats/free/FreeT.scala
+++ b/free/src/main/scala/cats/free/FreeT.scala
@@ -220,7 +220,12 @@ object FreeT extends FreeTInstances {
 }
 
 sealed abstract private[free] class FreeTInstances extends FreeTInstances0 {
-  implicit def catsFreeMonadErrorForFreeT[S[_], M[_], E](implicit E: MonadError[M, E]): MonadError[FreeT[S, M, *], E] =
+
+  // retained for binary compatibility. its results are incorrect though and it would fail the laws if we generated things of the form pure(()).flatMap(_ => fa)
+  @deprecated("does not handle errors beyond the head suspension; use catsFreeMonadErrorForFreeT2", "2.1.0")
+  def catsFreeMonadErrorForFreeT[S[_], M[_], E](
+    implicit E: MonadError[M, E]
+  ): MonadError[FreeT[S, M, *], E] =
     new MonadError[FreeT[S, M, *], E] with FreeTMonad[S, M] {
       override def M = E
       override def handleErrorWith[A](fa: FreeT[S, M, A])(f: E => FreeT[S, M, A]) =
@@ -228,6 +233,66 @@ sealed abstract private[free] class FreeTInstances extends FreeTInstances0 {
       override def raiseError[A](e: E) =
         FreeT.liftT(E.raiseError[A](e))(M)
     }
+
+  // not to be confused with defer... which is something different... sigh...
+  implicit def catsDeferForFreeT[S[_], M[_]: Applicative]: Defer[FreeT[S, M, *]] =
+    new Defer[FreeT[S, M, *]] {
+      def defer[A](fa: => FreeT[S, M, A]): FreeT[S, M, A] =
+        FreeT.pure[S, M, Unit](()).flatMap(_ => fa)
+    }
+
+  implicit def catsFreeMonadErrorForFreeT2[S[_], M[_], E](implicit E: MonadError[M, E],
+                                                          S: Functor[S]): MonadError[FreeT[S, M, *], E] =
+    new MonadError[FreeT[S, M, *], E] with FreeTMonad[S, M] {
+      override def M = E
+
+      private[this] val RealDefer = catsDeferForFreeT[S, M]
+
+      /*
+       * Quick explanation... The previous version of this function (retained above for
+       * bincompat) was only able to look at the *top* level M[_] suspension in a Free
+       * program. Any suspensions below that in the compute tree were invisible. Thus,
+       * if there were errors in that top level suspension, then they would be handled
+       * by the delegate. Errors buried further in the tree were unhandled. This is most
+       * easily visualized by the following two expressions:
+       *
+       * - fa
+       * - pure(()).flatMap(_ => fa)
+       *
+       * By the monad laws, these *should* be identical in effect, but they do have
+       * different structural representations within FreeT. More specifically, the latter
+       * has a meaningless M[_] suspension which sits in front of the rest of the
+       * computation. The previous iteration of this function would be blind to fa in
+       * the latter encoding, while it would handle it correctly in the former.
+       *
+       * Historical sidebar: this became visible because of the "shift" mechanism in
+       * Kleisli.
+       */
+      override def handleErrorWith[A](fa: FreeT[S, M, A])(f: E => FreeT[S, M, A]) = {
+        val ft = FreeT.liftT[S, M, FreeT[S, M, A]] {
+          val resultsM = E.map(fa.resume) {
+            case Left(se) =>
+              // we defer here in order to ensure stack-safety in the results even when M[_] is not itself stack-safe
+              // there's some small performance loss as a consequence, but really, if you care that much about performance, why are you using FreeT?
+              RealDefer.defer(FreeT.liftF[S, M, FreeT[S, M, A]](S.map(se)(handleErrorWith(_)(f))).flatMap(identity))
+
+            case Right(a) =>
+              pure(a)
+          }
+
+          E.handleErrorWith(resultsM) { e =>
+            E.map(f(e).resume) { eth =>
+              FreeT.defer(E.pure(eth.swap)) // why on earth is defer inconsistent with resume??
+            }
+          }
+        }
+
+        ft.flatMap(identity)
+      }
+
+      override def raiseError[A](e: E) =
+        FreeT.liftT(E.raiseError[A](e))(M)
+    }
 }
 
 sealed abstract private[free] class FreeTInstances0 extends FreeTInstances1 {
diff --git a/free/src/test/scala/cats/free/FreeTSuite.scala b/free/src/test/scala/cats/free/FreeTSuite.scala
index 5fcdd8f55a..2210a23816 100644
--- a/free/src/test/scala/cats/free/FreeTSuite.scala
+++ b/free/src/test/scala/cats/free/FreeTSuite.scala
@@ -44,6 +44,8 @@ class FreeTSuite extends CatsSuite {
              SerializableTests.serializable(MonadError[FreeTOption, Unit]))
   }
 
+  checkAll("FreeT[Option, Option, Int", DeferTests[FreeTOption].defer[Int])
+
   test("FlatMap stack safety tested with 50k flatMaps") {
     val expected = Applicative[FreeTOption].pure(())
     val result =
@@ -115,6 +117,15 @@ class FreeTSuite extends CatsSuite {
     }
   }
 
+  // NB: this does not analogously cause problems for the SemigroupK implementation as semigroup's effects associate while errors do not
+  test("handle errors in non-head suspensions") {
+    type F[A] = FreeT[Id, Option, A]
+    val F = MonadError[F, Unit]
+
+    val eff = F.flatMap(F.pure(()))(_ => F.raiseError[String](()))
+    F.attempt(eff).runM(Some(_)) should ===(Some(Left(())))
+  }
+
   sealed trait Test1Algebra[A]
 
   case class Test1[A](value: Int, f: Int => A) extends Test1Algebra[A]
diff --git a/laws/src/main/scala/cats/laws/ApplicativeErrorLaws.scala b/laws/src/main/scala/cats/laws/ApplicativeErrorLaws.scala
index 2b6c079a82..dd061468ee 100644
--- a/laws/src/main/scala/cats/laws/ApplicativeErrorLaws.scala
+++ b/laws/src/main/scala/cats/laws/ApplicativeErrorLaws.scala
@@ -54,6 +54,21 @@ trait ApplicativeErrorLaws[F[_], E] extends ApplicativeLaws[F] {
 
   def redeemDerivedFromAttemptMap[A, B](fa: F[A], fe: E => B, fs: A => B): IsEq[F[B]] =
     F.redeem(fa)(fe, fs) <-> F.map(F.attempt(fa))(_.fold(fe, fs))
+
+  /*
+   * These laws, taken together with applicativeErrorHandle, show that errors dominate in
+   * ap, *and* show that handle has lexical semantics over ap. F.unit is used in both laws
+   * because we don't have another way of expressing "an F[_] which does *not* contain any
+   * errors". We could make these laws considerably stronger if such a thing were
+   * expressible. Specifically, what we're missing here is the ability to say that
+   * raiseError distributes over an *arbitrary* number of aps.
+   */
+
+  def raiseErrorDistributesOverApLeft[A](h: E => F[A], e: E) =
+    F.handleErrorWith(F.ap(F.raiseError[Unit => A](e))(F.unit))(h) <-> h(e)
+
+  def raiseErrorDistributesOverApRight[A](h: E => F[A], e: E) =
+    F.handleErrorWith(F.ap(F.pure((a: A) => a))(F.raiseError[A](e)))(h) <-> h(e)
 }
 
 object ApplicativeErrorLaws {
diff --git a/laws/src/main/scala/cats/laws/discipline/ApplicativeErrorTests.scala b/laws/src/main/scala/cats/laws/discipline/ApplicativeErrorTests.scala
index 9e89895c9e..e53d3918fd 100644
--- a/laws/src/main/scala/cats/laws/discipline/ApplicativeErrorTests.scala
+++ b/laws/src/main/scala/cats/laws/discipline/ApplicativeErrorTests.scala
@@ -56,7 +56,13 @@ trait ApplicativeErrorTests[F[_], E] extends ApplicativeTests[F] {
         "applicativeError onError raise" -> forAll(laws.onErrorRaise[A] _),
         "applicativeError adaptError pure" -> forAll(laws.adaptErrorPure[A] _),
         "applicativeError adaptError raise" -> forAll(laws.adaptErrorRaise[A] _),
-        "applicativeError redeem is derived from attempt and map" -> forAll(laws.redeemDerivedFromAttemptMap[A, B] _)
+        "applicativeError redeem is derived from attempt and map" -> forAll(laws.redeemDerivedFromAttemptMap[A, B] _),
+        "applicativeError handleError . raiseError left-distributes over ap" -> forAll(
+          laws.raiseErrorDistributesOverApLeft[A] _
+        ),
+        "applicativeError handleError . raiseError right-distributes over ap" -> forAll(
+          laws.raiseErrorDistributesOverApRight[A] _
+        )
       )
     }
 }