diff --git a/src/google/adk/evaluation/agent_evaluator.py b/src/google/adk/evaluation/agent_evaluator.py
index f52a367950..8548093778 100644
--- a/src/google/adk/evaluation/agent_evaluator.py
+++ b/src/google/adk/evaluation/agent_evaluator.py
@@ -32,6 +32,7 @@
 from pydantic import ValidationError
 
 from ..agents.base_agent import BaseAgent
+from ..utils import json_utils
 from ..utils.context_utils import Aclosing
 from .constants import MISSING_EVAL_DEPENDENCIES_MESSAGE
 from .eval_case import get_all_tool_calls
@@ -324,7 +325,9 @@ def _get_initial_session(initial_session_file: Optional[str] = None):
     initial_session = {}
     if initial_session_file:
       with open(initial_session_file, "r") as f:
-        initial_session = json.loads(f.read())
+        initial_session = json_utils.safe_json_loads(
+            f.read(), context=initial_session_file
+        )
     return initial_session
 
   @staticmethod
diff --git a/src/google/adk/integrations/vmaas/sandbox_client.py b/src/google/adk/integrations/vmaas/sandbox_client.py
index 1a264a1146..84394b40d3 100644
--- a/src/google/adk/integrations/vmaas/sandbox_client.py
+++ b/src/google/adk/integrations/vmaas/sandbox_client.py
@@ -28,6 +28,7 @@
 
 from ...features import experimental
 from ...features import FeatureName
+from ...utils import json_utils
 
 if TYPE_CHECKING:
   import vertexai
@@ -129,10 +130,10 @@ def _parse_response(self, response: Any) -> dict[str, Any]:
     Returns:
       The parsed JSON response as a dict.
     """
-    import json
-
     if hasattr(response, "body") and response.body:
-      return json.loads(response.body)
+      return json_utils.safe_json_loads(
+          response.body, context="sandbox response"
+      )
     return {}
 
   def update_access_token(self, access_token: str) -> None:
diff --git a/src/google/adk/models/anthropic_llm.py b/src/google/adk/models/anthropic_llm.py
index 9658b85a5f..63f89ed6c8 100644
--- a/src/google/adk/models/anthropic_llm.py
+++ b/src/google/adk/models/anthropic_llm.py
@@ -41,6 +41,7 @@
 from pydantic import BaseModel
 from typing_extensions import override
 
+from ..utils import json_utils
 from ..utils._google_client_headers import get_tracking_headers
 from .base_llm import BaseLlm
 from .llm_response import LlmResponse
@@ -75,29 +76,20 @@ def _build_anthropic_thinking_param(
 ) -> Union[
     anthropic_types.ThinkingConfigEnabledParam,
     anthropic_types.ThinkingConfigDisabledParam,
-    anthropic_types.ThinkingConfigAdaptiveParam,
     NotGiven,
 ]:
   """Maps genai ThinkingConfig to Anthropic's thinking parameter.
 
   Per ``google.genai.types.ThinkingConfig``, ``thinking_budget`` semantics are:
     * ``None``: not specified; the genai default is model-dependent. Anthropic
-      requires an explicit choice whenever thinking is configured, so we
-      surface this as a ``ValueError`` to keep the developer's intent
+      requires an explicit ``budget_tokens`` whenever thinking is enabled, so
+      we surface this as a ``ValueError`` to keep the developer's intent
       explicit (mirroring the Anthropic API).
-    * ``0``: thinking is DISABLED (``thinking.type: "disabled"``).
-    * negative (e.g. ``-1`` AUTOMATIC): maps to Anthropic's adaptive thinking
-      (``thinking.type: "adaptive"``). The model picks the depth itself
-      (controlled by the separate ``output_config.effort`` parameter when
-      set). REQUIRED for Claude Opus 4.7 and later models that reject
-      ``"enabled"`` with a 400 error; also recommended for Opus 4.6 and
-      Sonnet 4.6 where ``"enabled"`` is deprecated.
-    * positive int: budget in tokens for legacy manual mode
-      (``thinking.type: "enabled"``; Anthropic requires ``>= 1024`` and
+    * ``0``: thinking is DISABLED.
+    * ``-1``: AUTOMATIC; not supported by Anthropic models.
+    * positive int: budget in tokens (Anthropic requires ``>= 1024`` and
       ``< max_tokens``; validation is delegated to the Anthropic API so the
-      caller gets the canonical error message). Rejected by Claude Opus 4.7
-      -- callers targeting 4.7+ must use a negative value (adaptive) or
-      ``0`` (disabled).
+      caller gets the canonical error message).
   """
   if not config or not config.thinking_config:
     return NOT_GIVEN
@@ -107,22 +99,19 @@ def _build_anthropic_thinking_param(
   if thinking_budget is None:
     raise ValueError(
         "thinking_budget must be set explicitly when ThinkingConfig is"
-        " provided for Anthropic models. Use 0 to disable thinking, -1 for"
-        " adaptive (model-chosen depth), or a positive integer (>= 1024)"
-        " for manual budgeting."
+        " provided for Anthropic models. Use 0 to disable thinking, or a"
+        " positive integer (>= 1024) for the token budget."
     )
 
   if thinking_budget == 0:
     return anthropic_types.ThinkingConfigDisabledParam(type="disabled")
 
   if thinking_budget < 0:
-    # genai AUTOMATIC (-1) and any other negative value map to Anthropic
-    # adaptive thinking. Required for Claude Opus 4.7 (which returns a 400
-    # error for ``"enabled"``) and recommended for Opus 4.6 / Sonnet 4.6
-    # where ``"enabled"`` is deprecated. Adaptive does not accept a budget;
-    # depth is controlled by the model itself (or by the separate
-    # ``output_config.effort`` parameter when set).
-    return anthropic_types.ThinkingConfigAdaptiveParam(type="adaptive")
+    raise ValueError(
+        f"thinking_budget={thinking_budget} is not supported for Anthropic"
+        " models (AUTOMATIC mode is unavailable). Use a positive integer"
+        " (>= 1024) for the token budget, or 0 to disable thinking."
+    )
 
   return anthropic_types.ThinkingConfigEnabledParam(
       type="enabled",
@@ -693,7 +682,9 @@ async def _generate_content_streaming(
         all_parts.append(types.Part.from_text(text=text_blocks[idx]))
       if idx in tool_use_blocks:
         acc = tool_use_blocks[idx]
-        args = json.loads(acc.args_json) if acc.args_json else {}
+        args = (
+            json_utils.safe_json_loads(acc.args_json) if acc.args_json else {}
+        )
         part = types.Part.from_function_call(name=acc.name, args=args)
         part.function_call.id = acc.id
         all_parts.append(part)
diff --git a/src/google/adk/models/apigee_llm.py b/src/google/adk/models/apigee_llm.py
index a1575bdce6..8966bfb221 100644
--- a/src/google/adk/models/apigee_llm.py
+++ b/src/google/adk/models/apigee_llm.py
@@ -35,6 +35,7 @@
 import tenacity
 from typing_extensions import override
 
+from ..utils import json_utils
 from ..utils.env_utils import is_env_enabled
 from .google_llm import Gemini
 from .llm_response import LlmResponse
@@ -572,7 +573,7 @@ async def _handle_streaming(
         try:
           for res in self._parse_streaming_line(line, accumulator):
             yield res
-        except json.JSONDecodeError:
+        except ValueError:
           logger.warning('Failed to parse JSON chunk: %s', line)
           continue
 
@@ -848,7 +849,7 @@ def _parse_streaming_line(
     Yields:
       An LlmResponse object parsed from the streaming line.
     """
-    chunk = json.loads(line)
+    chunk = json_utils.safe_json_loads(line, context='streaming response')
     for response in accumulator.process_chunk(chunk):
       yield response
 
@@ -1160,15 +1161,14 @@ def _upsert_tool_call(self, tool_call: dict[str, Any]) -> types.Part:
     func = tool_call.get('function', {})
     args_delta = func.get('arguments', '')
     if args_delta:
-      try:
-        args = json.loads(args_delta)
-        chunk_part.function_call.args = args
-        if not part.function_call.args:
-          part.function_call.args = dict(args)
-        else:
-          part.function_call.args.update(args)
-      except json.JSONDecodeError as e:
-        raise ValueError(f'Failed to parse arguments: {args_delta}') from e
+      args = json_utils.safe_json_loads(
+          args_delta, context='streaming response'
+      )
+      chunk_part.function_call.args = args
+      if not part.function_call.args:
+        part.function_call.args = dict(args)
+      else:
+        part.function_call.args.update(args)
 
     func_name = func.get('name')
     if func_name:
diff --git a/src/google/adk/sessions/schemas/shared.py b/src/google/adk/sessions/schemas/shared.py
index 25d4ea9e95..582c95c6d7 100644
--- a/src/google/adk/sessions/schemas/shared.py
+++ b/src/google/adk/sessions/schemas/shared.py
@@ -15,6 +15,7 @@
 
 import json
 
+from google.adk.utils import json_utils
 from sqlalchemy import Dialect
 from sqlalchemy import Text
 from sqlalchemy.dialects import mysql
@@ -51,7 +52,7 @@ def process_result_value(self, value, dialect: Dialect):
       if dialect.name == "postgresql":
         return value  # JSONB returns dict directly
       else:
-        return json.loads(value)  # Deserialize from JSON string for TEXT
+        return json_utils.safe_json_loads(value, context="session state")
     return value
 
 
diff --git a/src/google/adk/sessions/sqlite_session_service.py b/src/google/adk/sessions/sqlite_session_service.py
index 798befcedc..6ed19d7ccc 100644
--- a/src/google/adk/sessions/sqlite_session_service.py
+++ b/src/google/adk/sessions/sqlite_session_service.py
@@ -27,6 +27,7 @@
 import aiosqlite
 from google.adk.platform import time as platform_time
 from google.adk.platform import uuid as platform_uuid
+from google.adk.utils import json_utils
 from typing_extensions import override
 
 from . import _session_util
@@ -245,7 +246,9 @@ async def get_session(
         session_row = await cursor.fetchone()
         if session_row is None:
           return None
-        session_state = json.loads(session_row["state"])
+        session_state = json_utils.safe_json_loads(
+            session_row["state"], context="session state"
+        )
         last_update_time = session_row["update_time"]
 
       # Build events query
@@ -328,12 +331,16 @@ async def list_sessions(
             (app_name,),
         ) as cursor:
           async for row in cursor:
-            user_states_map[row["user_id"]] = json.loads(row["state"])
+            user_states_map[row["user_id"]] = json_utils.safe_json_loads(
+                row["state"], context="session state"
+            )
 
       # Build session list
       for row in session_rows:
         session_user_id = row["user_id"]
-        session_state = json.loads(row["state"])
+        session_state = json_utils.safe_json_loads(
+            row["state"], context="session state"
+        )
         user_state = user_states_map.get(session_user_id, {})
         merged_state = _merge_state(app_state, user_state, session_state)
         sessions_list.append(
@@ -391,7 +398,7 @@ async def append_event(self, session: Session, event: Event) -> Event:
 
       # Apply state delta if present
       has_session_state_delta = False
-      if event.actions.state_delta:
+      if event.actions and event.actions.state_delta:
         state_deltas = _session_util.extract_state_delta(
             event.actions.state_delta
         )
@@ -475,7 +482,11 @@ async def _get_state(
     """Fetches and deserializes a JSON state column from a single row."""
     async with db.execute(query, params) as cursor:
       row = await cursor.fetchone()
-      return json.loads(row["state"]) if row else {}
+      return (
+          json_utils.safe_json_loads(row["state"], context="session state")
+          if row
+          else {}
+      )
 
   async def _get_app_state(
       self, db: aiosqlite.Connection, app_name: str
diff --git a/src/google/adk/utils/_schema_utils.py b/src/google/adk/utils/_schema_utils.py
index e83431bd61..0bca75343f 100644
--- a/src/google/adk/utils/_schema_utils.py
+++ b/src/google/adk/utils/_schema_utils.py
@@ -30,6 +30,8 @@
 from pydantic import BaseModel
 from pydantic import TypeAdapter
 
+from . import json_utils
+
 # Use SchemaUnion from google.genai.types to support all schema types
 # that the underlying API supports.
 SchemaType = types.SchemaUnion
@@ -130,4 +132,4 @@ def validate_schema(schema: SchemaType, json_text: str) -> Any:
   else:
     # For other schema types (list[str], dict, Schema, etc.),
     # just parse JSON without pydantic validation
-    return json.loads(json_text)
+    return json_utils.safe_json_loads(json_text, context='schema value')
diff --git a/src/google/adk/utils/json_utils.py b/src/google/adk/utils/json_utils.py
new file mode 100644
index 0000000000..b0a4c57068
--- /dev/null
+++ b/src/google/adk/utils/json_utils.py
@@ -0,0 +1,46 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import json
+from typing import Any
+from typing import Optional
+
+
+def safe_json_loads(text: str, context: Optional[str] = None) -> Any:
+  """Parses a JSON string, raising ValueError on malformed input.
+
+  Wraps ``json.loads`` with a consistent error type so callers don't need
+  to handle ``json.JSONDecodeError`` directly. All JSON parsing in the
+  ADK runtime should go through this helper so errors surface with a
+  clear, actionable message.
+
+  Args:
+    text: The JSON string to parse.
+    context: Optional human-readable label for the source of ``text``
+      (e.g. ``"session state"``), included in the error message to aid
+      debugging.
+
+  Returns:
+    The parsed Python object.
+
+  Raises:
+    ValueError: If ``text`` is not valid JSON.
+  """
+  try:
+    return json.loads(text)
+  except json.JSONDecodeError as exc:
+    suffix = f' in {context}' if context else ''
+    raise ValueError(f'Invalid JSON{suffix}: {exc}') from exc
diff --git a/tests/unittests/utils/test_json_utils.py b/tests/unittests/utils/test_json_utils.py
new file mode 100644
index 0000000000..9e256764d8
--- /dev/null
+++ b/tests/unittests/utils/test_json_utils.py
@@ -0,0 +1,90 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for JSON utility functions."""
+
+from google.adk.utils.json_utils import safe_json_loads
+import pytest
+
+
+def test_parses_object():
+  result = safe_json_loads('{"key": "value"}')
+  assert result == {'key': 'value'}
+
+
+def test_parses_array():
+  result = safe_json_loads('[1, 2, 3]')
+  assert result == [1, 2, 3]
+
+
+def test_parses_nested_structure():
+  result = safe_json_loads('{"a": {"b": [1, null, true]}}')
+  assert result == {'a': {'b': [1, None, True]}}
+
+
+def test_parses_string_value():
+  result = safe_json_loads('"hello"')
+  assert result == 'hello'
+
+
+def test_parses_number():
+  result = safe_json_loads('42')
+  assert result == 42
+
+
+def test_parses_null():
+  result = safe_json_loads('null')
+  assert result is None
+
+
+def test_malformed_raises_value_error():
+  with pytest.raises(ValueError):
+    safe_json_loads('{bad json}')
+
+
+def test_empty_string_raises_value_error():
+  with pytest.raises(ValueError):
+    safe_json_loads('')
+
+
+def test_error_message_includes_context():
+  with pytest.raises(ValueError, match='session state'):
+    safe_json_loads('{bad}', context='session state')
+
+
+def test_error_message_without_context():
+  with pytest.raises(ValueError, match='Invalid JSON'):
+    safe_json_loads('{bad}')
+
+
+def test_error_wraps_json_decode_error():
+  with pytest.raises(ValueError) as exc_info:
+    safe_json_loads('{bad}', context='test')
+  assert exc_info.value.__cause__ is not None
+  import json
+
+  assert isinstance(exc_info.value.__cause__, json.JSONDecodeError)
+
+
+def test_context_none_no_suffix():
+  with pytest.raises(ValueError) as exc_info:
+    safe_json_loads('{bad}', context=None)
+  msg = str(exc_info.value)
+  assert msg.startswith('Invalid JSON:'), msg
+  assert not msg.startswith('Invalid JSON in '), msg
+
+
+def test_unicode_content():
+  result = safe_json_loads('{"emoji": "🎉", "chinese": "你好"}')
+  assert result == {'emoji': '🎉', 'chinese': '你好'}