diff --git a/docs/adr/0004-use-handle-to-model-effects.md b/docs/adr/0004-use-handle-to-model-effects.md
new file mode 100644
index 00000000000..a0034edacef
--- /dev/null
+++ b/docs/adr/0004-use-handle-to-model-effects.md
@@ -0,0 +1,37 @@
+# 4. Use Handle to model Effects
+
+Date: 2021-06-08
+
+## Status
+
+Accepted
+
+## Context
+
+* Given we are structuring Hydra node as a [reactive core](0002-reactive-core) we need a way to ensure a strict separation of pure and impure (or effectful) code
+* Various _patterns_ have been proposed in Haskell to model effects:
+  * [Tagless final encoding](http://okmij.org/ftp/tagless-final/index.html) also known as _MTL-style_ although using typeclasses to implement is [not necessary](https://www.foxhound.systems/blog/final-tagless/), whereby Effect(s) are expressed as typeclass(es) which are propagated as constraints
+  * [Free monads](https://reasonablypolymorphic.com/blog/freer-monads/), or any variant thereof like Eff, freer, extensible-effects, whereby effect(s) are expressed as ADTs which are _interpreted_ in the context of an _Effect stack_
+  * [Handle](https://jaspervdj.be/posts/2018-03-08-handle-pattern.html) pattern also known as _record-of-functions_ whereby effects are grouped together in a datatype with a single record constructor
+* All options considered have both advocates and detractors and there is no widely accepted pattern
+* We did a couple experiments to explore various options
+
+## Decision
+
+* Effectful components of the Hydra node will be defined using  the _Handle pattern_
+
+## Consequences
+
+### Effects
+
+* For example, the on-chain component is defined as:
+ ```
+  newtype OnChain m = OnChain
+  { postTx :: MonadThrow m => OnChainTx -> m ()
+  }
+  ```
+* Instantiation to concrete IO is pushed at the outermost layer, eg. in the `Main` or tests
+
+## Additional comments
+
+* These tradeoffs also appear in other functional languages like [F#](https://medium.com/@dogwith1eye/prefer-records-of-functions-to-interfaces-d6413af4d2c3)