diff --git a/pydantic_evals/pydantic_evals/otel/_context_in_memory_span_exporter.py b/pydantic_evals/pydantic_evals/otel/_context_in_memory_span_exporter.py
index cd83f30a99..05d4d6bff9 100644
--- a/pydantic_evals/pydantic_evals/otel/_context_in_memory_span_exporter.py
+++ b/pydantic_evals/pydantic_evals/otel/_context_in_memory_span_exporter.py
@@ -109,7 +109,7 @@ def get_finished_spans(self, context_id: str | None = None) -> tuple[ReadableSpa
 
     def export(self, spans: typing.Sequence[ReadableSpan]) -> SpanExportResult:
         """Stores a list of spans in memory."""
-        if self._stopped:  # pragma: no cover
+        if self._stopped:
             return SpanExportResult.FAILURE
         with self._lock:
             context_id = _EXPORTER_CONTEXT_ID.get()
diff --git a/tests/conftest.py b/tests/conftest.py
index f37d35a946..26a6ad039d 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -24,6 +24,7 @@
 from vcr import VCR, request as vcr_request
 
 import pydantic_ai.models
+from pydantic_ai import Agent
 from pydantic_ai.messages import BinaryContent
 from pydantic_ai.models import Model, cached_async_http_client
 
@@ -230,6 +231,24 @@ def event_loop() -> Iterator[None]:
     new_loop.close()
 
 
+@pytest.fixture(autouse=True)
+def no_instrumentation_by_default():
+    Agent.instrument_all(False)
+
+
+try:
+    import logfire
+
+    logfire.DEFAULT_LOGFIRE_INSTANCE.config.ignore_no_config = True
+
+    @pytest.fixture(autouse=True)
+    def fresh_logfire():
+        logfire.shutdown(flush=False)
+
+except ImportError:
+    pass
+
+
 def raise_if_exception(e: Any) -> None:
     if isinstance(e, Exception):
         raise e
diff --git a/tests/evals/test_otel.py b/tests/evals/test_otel.py
index 10fec9f7bb..715742735e 100644
--- a/tests/evals/test_otel.py
+++ b/tests/evals/test_otel.py
@@ -336,9 +336,6 @@ async def test_span_node_repr(span_tree: SpanTree):
 
 async def test_span_tree_ancestors_methods():
     """Test the ancestor traversal methods in SpanNode."""
-    # Configure logfire
-    logfire.configure()
-
     # Create spans with a deep structure for testing ancestor methods
     with context_subtree() as tree:
         with logfire.span('root', depth=0):
@@ -398,9 +395,6 @@ async def test_span_tree_ancestors_methods():
 
 async def test_span_tree_descendants_methods():
     """Test the descendant traversal methods in SpanNode."""
-    # Configure logfire
-    logfire.configure()
-
     # Create spans with a deep structure for testing descendant methods
     with context_subtree() as tree:
         with logfire.span('root', depth=0):
@@ -488,9 +482,6 @@ async def test_span_tree_descendants_methods():
 
 async def test_log_levels_and_exceptions():
     """Test recording different log levels and exceptions in spans."""
-    # Configure logfire
-    logfire.configure()
-
     with context_subtree() as tree:
         # Test different log levels
         with logfire.span('parent_span'):
@@ -891,7 +882,6 @@ async def test_context_subtree_not_configured(mocker: MockerFixture):
     """Test that context_subtree correctly records spans in independent async contexts."""
     from opentelemetry.trace import ProxyTracerProvider
 
-    # from opentelemetry.sdk.trace import TracerProvider
     mocker.patch(
         'pydantic_evals.otel._context_in_memory_span_exporter.get_tracer_provider', return_value=ProxyTracerProvider()
     )