diff --git a/tests/experiments/test_loop.py b/tests/experiments/test_loop.py
new file mode 100644
index 000000000..7fcb1c944
--- /dev/null
+++ b/tests/experiments/test_loop.py
@@ -0,0 +1,126 @@
+import gzip
+import json
+import pickle
+
+import numpy as np
+import pytest
+
+from agentlab.experiments.loop import DataclassJSONEncoder, ExpResult, StepInfo, StepTimestamps
+
+GOAL_OBJECT = [{"type": "text", "text": "do the thing"}]
+
+
+def _make_step_info(step=0):
+    rng = np.random.default_rng(0)
+    screenshot = rng.integers(0, 256, size=(4, 5, 3), dtype=np.uint8)
+    screenshot_som = rng.integers(0, 256, size=(4, 5, 3), dtype=np.uint8)
+    obs = {
+        "screenshot": screenshot,
+        "screenshot_som": screenshot_som,
+        "goal_object": list(GOAL_OBJECT),
+        "axtree_txt": "RootWebArea",
+    }
+    return StepInfo(
+        step=step,
+        obs=obs,
+        reward=np.float32(0.5),
+        raw_reward=np.int64(1),
+        terminated=False,
+        truncated=False,
+        action="click('12')",
+        agent_info={"think": "some thought"},
+        stats={"n_token_axtree_txt": np.int64(3)},
+        profiling=StepTimestamps(env_start=1.0, env_stop=2.0),
+    )
+
+
+def test_save_step_info_round_trip(tmp_path):
+    step_info = _make_step_info()
+    screenshot = step_info.obs["screenshot"]
+    screenshot_som = step_info.obs["screenshot_som"]
+    goal_object = step_info.obs["goal_object"]
+
+    step_info.save_step_info(tmp_path, save_screenshot=True, save_som=True)
+
+    assert (tmp_path / "screenshot_step_0.png").exists()
+    assert (tmp_path / "screenshot_som_step_0.png").exists()
+    assert (tmp_path / "goal_object.pkl.gz").exists()
+
+    # obs is restored in memory, except for the offloaded goal object
+    np.testing.assert_array_equal(step_info.obs["screenshot"], screenshot)
+    np.testing.assert_array_equal(step_info.obs["screenshot_som"], screenshot_som)
+    assert step_info.obs["goal_object"] is None
+
+    # the pickled step contains neither the screenshots nor the goal object
+    with gzip.open(tmp_path / "step_0.pkl.gz", "rb") as f:
+        raw_step = pickle.load(f)
+    assert "screenshot" not in raw_step.obs
+    assert "screenshot_som" not in raw_step.obs
+    assert raw_step.obs["goal_object"] is None
+
+    loaded = ExpResult(tmp_path).get_step_info(0)
+    np.testing.assert_array_equal(loaded.obs["screenshot"], screenshot)
+    np.testing.assert_array_equal(loaded.obs["screenshot_som"], screenshot_som)
+    assert loaded.obs["goal_object"] == goal_object
+    assert loaded.action == step_info.action
+    assert loaded.obs["axtree_txt"] == step_info.obs["axtree_txt"]
+
+
+def test_save_step_info_goal_object_saved_once(tmp_path):
+    first = _make_step_info(step=0)
+    first.save_step_info(tmp_path)
+
+    second = _make_step_info(step=1)
+    second.obs["goal_object"] = [{"type": "text", "text": "a different goal"}]
+    second.save_step_info(tmp_path)
+
+    with gzip.open(tmp_path / "goal_object.pkl.gz", "rb") as f:
+        assert pickle.load(f) == GOAL_OBJECT
+
+    assert ExpResult(tmp_path).get_step_info(1).obs["goal_object"] == GOAL_OBJECT
+
+
+def test_save_step_info_without_screenshots(tmp_path):
+    step_info = _make_step_info()
+    step_info.save_step_info(tmp_path, save_screenshot=False, save_som=False)
+
+    assert not (tmp_path / "screenshot_step_0.png").exists()
+    assert not (tmp_path / "screenshot_som_step_0.png").exists()
+
+    loaded = ExpResult(tmp_path).get_step_info(0)
+    assert "screenshot" not in loaded.obs
+    assert "screenshot_som" not in loaded.obs
+
+
+def test_save_step_info_json(tmp_path):
+    step_info = _make_step_info()
+    step_info.obs["numpy_array"] = np.arange(3, dtype=np.int64)
+    step_info.save_step_info(tmp_path, save_json=True)
+
+    with open(tmp_path / "steps_info.json", "r") as f:
+        data = json.load(f)
+
+    assert data["step"] == 0
+    assert data["reward"] == pytest.approx(0.5)
+    assert data["raw_reward"] == 1
+    assert data["stats"]["n_token_axtree_txt"] == 3
+    assert data["obs"]["numpy_array"] == [0, 1, 2]
+    assert data["obs"]["goal_object"] is None
+
+
+def test_dataclass_json_encoder_numpy():
+    encoded = json.dumps(
+        {
+            "int": np.int64(3),
+            "float": np.float32(1.5),
+            "array": np.arange(2),
+            "dataclass": StepTimestamps(env_start=1.0),
+        },
+        cls=DataclassJSONEncoder,
+    )
+    decoded = json.loads(encoded)
+
+    assert decoded["int"] == 3
+    assert decoded["float"] == pytest.approx(1.5)
+    assert decoded["array"] == [0, 1]
+    assert decoded["dataclass"]["env_start"] == 1.0