diff --git a/velvetflow/bindings.py b/velvetflow/bindings.py
index 72ce852e..36e5f2cf 100644
--- a/velvetflow/bindings.py
+++ b/velvetflow/bindings.py
@@ -1037,7 +1037,14 @@ def _resolve(value: Any, path: str = "") -> Any:
             if stripped_template.startswith("{{") and stripped_template.endswith("}}"):
                 inner = stripped_template[2:-2].strip()
                 try:
-                    return eval_jinja_expression(inner, ctx.build_jinja_context())
+                    jinja_result = eval_jinja_expression(inner, ctx.build_jinja_context())
+                    try:
+                        from jinja2.runtime import Undefined  # type: ignore[import]
+                    except Exception:  # pragma: no cover - fallback when jinja is absent
+                        Undefined = ()  # type: ignore[assignment]
+
+                    if jinja_result is not None and not isinstance(jinja_result, Undefined):
+                        return jinja_result
                 except Exception:
                     pass
             rendered_inline = _render_json_bindings(normalized_templates)
diff --git a/velvetflow/executor/loops.py b/velvetflow/executor/loops.py
index fe3d1f14..5ffa73fa 100644
--- a/velvetflow/executor/loops.py
+++ b/velvetflow/executor/loops.py
@@ -33,10 +33,13 @@ def _apply_loop_exports(
         aggregates_output: Dict[str, Any],
         avg_state: Dict[str, Dict[str, float]],
         default_from_node: Optional[str],
-        extra_exports: Optional[Mapping[str, Any]],
-        extra_output: Dict[str, Any],
-        loop_binding: Optional[BindingContext],
+        extra_exports: Optional[Mapping[str, Any]] = None,
+        extra_output: Optional[Dict[str, Any]] = None,
+        loop_binding: Optional[BindingContext] = None,
     ) -> None:
+        if extra_output is None:
+            extra_output = {}
+
         normalized_items_spec: Optional[Mapping[str, Any]] = None
 
         if isinstance(items_spec, list):
diff --git a/velvetflow/planner/agent_runtime.py b/velvetflow/planner/agent_runtime.py
index 1fa39d4b..d92af37f 100644
--- a/velvetflow/planner/agent_runtime.py
+++ b/velvetflow/planner/agent_runtime.py
@@ -23,4 +23,51 @@
 from openai import OpenAI
 from pydantic import create_model
 
-from agents import Agent, Runner, function_tool
+try:  # Prefer the official Agents SDK when available
+    from agents import Agent, Runner, function_tool
+except Exception as import_error:  # pragma: no cover - exercised in CI without agents
+    import functools
+    import warnings
+
+    warnings.warn(
+        "Falling back to the built-in Agents shim because the optional "
+        "'agents' package is missing or incompatible. Install the official "
+        "package alongside openai>=1.30.0 to enable full agent execution.",
+        RuntimeWarning,
+    )
+
+    class Agent:
+        def __init__(self, *, name: str, instructions: str, tools: Sequence[Callable[..., Any]], model: str, **_: Any):
+            self.name = name
+            self.instructions = instructions
+            self.tools = list(tools)
+            self.model = model
+
+    def function_tool(strict_mode: bool = True, **decorator_kwargs: Any) -> Callable[[Callable[..., Any]], Callable[..., Any]]:
+        def _decorator(func: Callable[..., Any]) -> Callable[..., Any]:
+            @functools.wraps(func)
+            def _wrapper(*args: Any, **kwargs: Any) -> Any:
+                return func(*args, **kwargs)
+
+            # Preserve metadata that downstream code may introspect
+            _wrapper.__original_function__ = func  # type: ignore[attr-defined]
+            _wrapper.__function_tool_strict_mode__ = strict_mode  # type: ignore[attr-defined]
+            _wrapper.__function_tool_kwargs__ = decorator_kwargs  # type: ignore[attr-defined]
+            return _wrapper
+
+        return _decorator
+
+    class Runner:
+        @staticmethod
+        def run_sync(agent: Agent, run_input: Any, *, max_turns: int | None = None) -> None:
+            raise RuntimeError(
+                "agents shim does not implement execution. Install the 'agents' package "
+                "compatible with openai>=1.30.0 to run agents."
+            ) from import_error
+
+        @staticmethod
+        async def run(agent: Agent, run_input: Any, *, max_turns: int | None = None) -> None:
+            raise RuntimeError(
+                "agents shim does not implement execution. Install the 'agents' package "
+                "compatible with openai>=1.30.0 to run agents."
+            ) from import_error