diff --git a/test/libs/test_multiagent.py b/test/libs/test_multiagent.py
index 620f5a16eeb..e8a7bfe9c56 100644
--- a/test/libs/test_multiagent.py
+++ b/test/libs/test_multiagent.py
@@ -19,7 +19,11 @@
 from torchrl.collectors import Collector
 from torchrl.data import LazyTensorStorage, TensorDictReplayBuffer
 from torchrl.envs.libs.meltingpot import MeltingpotEnv, MeltingpotWrapper
-from torchrl.envs.libs.pettingzoo import _has_pettingzoo, PettingZooEnv
+from torchrl.envs.libs.pettingzoo import (
+    _has_pettingzoo,
+    PettingZooEnv,
+    PettingZooWrapper,
+)
 from torchrl.envs.libs.smacv2 import _has_smacv2, SMACv2Env
 from torchrl.envs.transforms import ActionMask, TransformedEnv
 from torchrl.envs.utils import check_env_specs, MarlGroupMapType
@@ -109,6 +113,90 @@ def test_dead_agents_done(self, seed=0):
                 ~mask
             ].all()  # When mask is false (dead agent), all agents are done
 
+    def test_action_mask_parallel_dead_agents(self):
+        """Regression test for #3702.
+
+        Parallel env + per-agent action mask + ``done_on_any=False``: when an
+        agent is removed from ``env.agents`` mid-episode, PettingZoo also drops
+        it from the observation dict, so ``_update_action_mask`` must skip
+        agents missing from ``observation_dict`` instead of indexing it blindly.
+        No stock PettingZoo env exercises this combination, so we use a tiny
+        custom ``ParallelEnv``.
+        """
+        import numpy as np
+        from gymnasium import spaces
+        from pettingzoo.utils.env import ParallelEnv
+
+        class _MaskedParallelEnv(ParallelEnv):
+            metadata = {"name": "masked_parallel_v0", "is_parallelizable": True}
+
+            def __init__(self):
+                self.possible_agents = ["a_0", "a_1", "a_2"]
+                self.agents = list(self.possible_agents)
+                self._n_actions = 2
+                self._obs_space = spaces.Dict(
+                    {
+                        "observation": spaces.Box(
+                            low=0.0, high=1.0, shape=(3,), dtype=np.float32
+                        ),
+                        "action_mask": spaces.MultiBinary(self._n_actions),
+                    }
+                )
+                self._act_space = spaces.Discrete(self._n_actions)
+                self._step_count = 0
+
+            def observation_space(self, agent):
+                return self._obs_space
+
+            def action_space(self, agent):
+                return self._act_space
+
+            def _agent_obs(self):
+                return {
+                    "observation": np.zeros(3, dtype=np.float32),
+                    "action_mask": np.ones(self._n_actions, dtype=np.int8),
+                }
+
+            def reset(self, seed=None, options=None):
+                self.agents = list(self.possible_agents)
+                self._step_count = 0
+                obs = {a: self._agent_obs() for a in self.agents}
+                infos = {a: {} for a in self.agents}
+                return obs, infos
+
+            def step(self, actions):
+                self._step_count += 1
+                # Drop a_0 starting from the 2nd step — mirrors PettingZoo's
+                # behavior when an agent terminates: it disappears from
+                # self.agents and from every per-agent dict returned by step.
+                if self._step_count >= 2 and "a_0" in self.agents:
+                    self.agents.remove("a_0")
+                obs = {a: self._agent_obs() for a in self.agents}
+                rewards = {a: 0.0 for a in self.agents}
+                terms = {a: False for a in self.agents}
+                truncs = {a: False for a in self.agents}
+                infos = {a: {} for a in self.agents}
+                return obs, rewards, terms, truncs, infos
+
+        env = PettingZooWrapper(
+            env=_MaskedParallelEnv(),
+            use_mask=True,
+            done_on_any=False,
+            seed=0,
+        )
+        try:
+            assert env.has_action_mask["a"]
+            # Pre-fix this rollout raises KeyError inside _update_action_mask
+            # the first time a_0 is missing from observation_dict.
+            td = env.rollout(max_steps=3, break_when_any_done=False)
+            # a_0 (index 0) is alive in step 1, dead in steps 2 and 3.
+            assert td["next", "a", "mask"][0, 0]
+            assert not td["next", "a", "mask"][1:, 0].any()
+            # The two surviving agents are alive throughout.
+            assert td["next", "a", "mask"][:, 1:].all()
+        finally:
+            env.close()
+
     @pytest.mark.parametrize(
         "wins_player_0",
         [True, False],
diff --git a/test/objectives/test_dqn.py b/test/objectives/test_dqn.py
index 7dcbac2dea0..433c0ba0507 100644
--- a/test/objectives/test_dqn.py
+++ b/test/objectives/test_dqn.py
@@ -745,7 +745,7 @@ def test_dqn_prioritized_weights(self):
         # Sample again - weights should now be non-equal
         sample2 = rb.sample()
         weights2 = sample2["priority_weight"]
-        assert weights2.std() > 1e-5
+        assert weights2.std() > 1e-6
 
         # Run loss again with varied weights
         loss_out2 = loss_fn(sample2)
@@ -1300,7 +1300,7 @@ def test_dqn_prioritized_weights(self):
         # Sample again - weights should now be non-equal
         sample2 = rb.sample()
         weights2 = sample2["priority_weight"]
-        assert weights2.std() > 1e-5
+        assert weights2.std() > 1e-6
 
         # Run loss again with varied weights
         loss_out2 = loss_fn(sample2)
diff --git a/torchrl/envs/libs/pettingzoo.py b/torchrl/envs/libs/pettingzoo.py
index 29f4febd0b5..2ab4521034d 100644
--- a/torchrl/envs/libs/pettingzoo.py
+++ b/torchrl/envs/libs/pettingzoo.py
@@ -804,6 +804,8 @@ def _update_action_mask(self, td, observation_dict, info_dict):
                 group_mask = td.get((group, "action_mask"))
                 group_mask += True
                 for index, agent in enumerate(agents):
+                    if agent not in observation_dict:
+                        continue
                     agent_obs = observation_dict[agent]
                     agent_info = info_dict[agent]
                     if isinstance(agent_obs, dict) and "action_mask" in agent_obs: