diff --git a/tests/test_agent_loop.py b/tests/test_agent_loop.py
new file mode 100644
index 0000000..7c3d75e
--- /dev/null
+++ b/tests/test_agent_loop.py
@@ -0,0 +1,105 @@
+"""Tests for the conversational LabAgent wrapper (lab_harness.agent.loop)."""
+
+from __future__ import annotations
+
+import pytest
+
+import lab_harness.agent.loop as loop_mod
+from lab_harness.agent.loop import LabAgent
+from lab_harness.config import ModelConfig, Settings
+from lab_harness.harness.engine.events import (
+    ErrorEvent,
+    StatusUpdate,
+    TextDelta,
+    ToolComplete,
+    ToolStart,
+    TurnComplete,
+)
+
+
+class _BrokenStore:
+    def __init__(self, *args, **kwargs):
+        raise RuntimeError("no memory backend")
+
+
+@pytest.fixture
+def settings() -> Settings:
+    return Settings(model=ModelConfig(provider="openai", model="gpt-test"))
+
+
+@pytest.fixture
+def agent(settings, monkeypatch) -> LabAgent:
+    # Keep tests hermetic: memory store construction fails, so no db file is created.
+    monkeypatch.setattr("lab_harness.memory.store.MemoryStore", _BrokenStore)
+    return LabAgent(settings=settings)
+
+
+def _fake_query(events, captured=None):
+    async def run_query(context, messages):
+        if captured is not None:
+            captured.append((context, messages))
+        for event in events:
+            yield event
+
+    return run_query
+
+
+def test_init_wires_registry_and_model_config(agent, settings):
+    assert agent.context.tool_registry is agent.registry
+    assert agent.context.model_config is settings.model
+
+
+def test_init_swallows_memory_store_failure(agent):
+    assert agent.context.memory_store is None
+
+
+def test_init_attaches_memory_store_when_available(settings, monkeypatch, tmp_path):
+    monkeypatch.chdir(tmp_path)
+    agent = LabAgent(settings=settings)
+    assert agent.context.memory_store is not None
+    assert (tmp_path / "data" / "memory.db").exists()
+
+
+async def test_run_conversation_joins_text_deltas_and_final_text(agent, monkeypatch):
+    captured = []
+    events = [TextDelta("Hello"), TextDelta(", "), TurnComplete(text="world")]
+    monkeypatch.setattr(loop_mod, "run_query", _fake_query(events, captured))
+
+    result = await agent.run_conversation("hi there")
+
+    assert result == "Hello, world"
+    context, messages = captured[0]
+    assert context is agent.context
+    assert messages == [{"role": "user", "content": "hi there"}]
+
+
+async def test_run_conversation_skips_empty_turn_complete(agent, monkeypatch):
+    events = [TextDelta("only delta"), TurnComplete(text="")]
+    monkeypatch.setattr(loop_mod, "run_query", _fake_query(events))
+
+    assert await agent.run_conversation("q") == "only delta"
+
+
+async def test_run_conversation_formats_errors(agent, monkeypatch):
+    events = [TextDelta("partial "), ErrorEvent(message="rate limited")]
+    monkeypatch.setattr(loop_mod, "run_query", _fake_query(events))
+
+    assert await agent.run_conversation("q") == "partial Error: rate limited"
+
+
+async def test_run_conversation_ignores_non_text_events(agent, monkeypatch):
+    events = [
+        StatusUpdate(message="thinking"),
+        ToolStart(tool_name="scan", tool_input={}),
+        ToolComplete(tool_name="scan", output="GPIB0::24"),
+        TurnComplete(text="done"),
+    ]
+    monkeypatch.setattr(loop_mod, "run_query", _fake_query(events))
+
+    assert await agent.run_conversation("q") == "done"
+
+
+async def test_run_conversation_with_no_events_returns_empty(agent, monkeypatch):
+    monkeypatch.setattr(loop_mod, "run_query", _fake_query([]))
+
+    assert await agent.run_conversation("q") == ""