diff --git a/pandag/pandag.py b/pandag/pandag.py
index 269a630..577a5d5 100644
--- a/pandag/pandag.py
+++ b/pandag/pandag.py
@@ -5,6 +5,8 @@
 from pandag.nodes import Node, Output
 from pandag import plot, graphml
 import more_itertools
+import itertools
+import logging
 
 
 class FakeDiGraph(nx.DiGraph):
@@ -127,6 +129,14 @@ def create_graph(self, sub, parent=None, local_dict=None, global_dict=None):
                                   local_dict=local_dict,
                                   global_dict=global_dict)
 
+    def start_nodes(self):
+        """Return nodes which don't have incoming edges."""
+        return [node for node in self.G.nodes if self.G.in_degree(node) == 0]
+
+    def end_nodes(self):
+        """Return nodes which don't have outgoing edges."""
+        return [node for node in self.G.nodes if self.G.out_degree(node) == 0]
+
     def eval(self, df):
         """Evaluate a Pandas DataFrame with the graph.
 
@@ -139,29 +149,56 @@ def eval(self, df):
         """
         # generate a unique column name
         node_col = f'{self.uuid}_curr_node'
-        for src_node_id, dst_node_id in nx.edge_dfs(self.G):
-            if node_col not in df.columns:
-                # this is the first node we're visiting, add its id to the
-                # `node_col` column as the current node ID for each rows
-                df[node_col] = src_node_id
-
-                # initialize the path column with the first node id as string, so
-                # we can later append new node IDs with a vectorized operation
-                if self.path_column is not None:
-                    df[self.path_column] = str(src_node_id)
-            edge_data = self.G.get_edge_data(src_node_id, dst_node_id)
-            src_node = self.get_node(src_node_id)
-            dst_node = self.get_node(dst_node_id)
-            if isinstance(dst_node, Output):
-                dst_node.update(df, df[node_col] == src_node_id)
-            flt = (df[node_col] == src_node_id) & (src_node.eval(df, edge_data))
+        # while not recommended, handle multiple start nodes (even multiple
+        # DAGs) in a general way, so we detect all start and end nodes
+        start_nodes_visited = set()
+        start_nodes = set(self.start_nodes())
+        end_nodes = set(self.end_nodes())
+        if len(start_nodes) > 1:
+            logging.warning(f"The DAG has {len(start_nodes)}, output might be non-deterministic!")
+
+        # loop through all start->end permutations
+        for start, end in itertools.product(start_nodes, end_nodes):
+            if start in end_nodes or end in start_nodes:
+                # exclude backwards permutations (end -> start)
+                continue
             if self.path_column:
-                # store the path which touched these rows
-                df.loc[flt, self.path_column] = df[self.path_column] + f',{dst_node_id}'
-            # update the matching rows' curr_node column to the next node,
-            # we'll use this to select the source rows for running the
-            # edge pointing from this node to the next
-            df.loc[flt, node_col] = dst_node_id
+                # for the first run, we initialize the column with a string,
+                # after that we're just appending
+                if start_nodes_visited:
+                    if start not in start_nodes_visited:
+                        df[self.path_column] = df[self.path_column] + f',{start}'
+                else:
+                    # initialize the path column with the first node id as string, so
+                    # we can later append new node IDs with a vectorized operation
+                    df[self.path_column] = str(start)
+            # (re)set the current node ID for each new start nodes
+            # if we've already visited this start node, leave the field alone,
+            # so we can make progress on sub-frames which are at a given node
+            # ID from a different path
+            if start not in start_nodes_visited:
+                df[node_col] = start
+            for path in nx.all_simple_edge_paths(self.G, start, end):
+                # this represents an edge between two nodes (src -> dst) in the path
+                for (src_node_id, dst_node_id) in path:
+                    edge_data = self.G.get_edge_data(src_node_id, dst_node_id)
+                    src_node = self.get_node(src_node_id)
+
+                    flt = (df[node_col] == src_node_id)
+                    if isinstance(src_node, Output):
+                        src_node.update(df, flt)
+                    else:
+                        flt &= (src_node.eval(df, edge_data))
+                    # update the matching rows' curr_node column to the next node,
+                    # we'll use this to select the source rows for running the
+                    # edge pointing from this node to the next
+                    df.loc[flt, node_col] = dst_node_id
+                    if self.path_column:
+                        # store the path which touched these rows
+                        df.loc[flt, self.path_column] = df[self.path_column] + f',{dst_node_id}'
+
+            # record that we've already visited this start node
+            start_nodes_visited.add(start)
 
         # remove the temporary current node column
         df = df.drop(node_col, axis=1)