diff --git a/docs/MCP_AND_TOOLS.md b/docs/MCP_AND_TOOLS.md
index 737289cf..627fd263 100644
--- a/docs/MCP_AND_TOOLS.md
+++ b/docs/MCP_AND_TOOLS.md
@@ -149,6 +149,8 @@ def book_flight(
 
 `agentctl validate` warns about any tool parameter without a description (or
 with a blank one). It does not fail validation yet.
+Parameters LangChain fills in at runtime, such as `config: RunnableConfig` or
+`InjectedToolArg` annotations, are never sent to the model and are not checked.
 
 Run `agentctl generate` to create tool stubs. The engine loads each tool once at
 graph-build time and wraps it as a LangChain `StructuredTool`. At runtime, only
diff --git a/docs/mcp-and-tools.mdx b/docs/mcp-and-tools.mdx
index 69465c25..edc58690 100644
--- a/docs/mcp-and-tools.mdx
+++ b/docs/mcp-and-tools.mdx
@@ -163,7 +163,7 @@ def get_order_status(
     ...
 ```
 
-`agentctl validate` warns about any tool parameter without a description, or with a blank one. It does not fail validation yet.
+`agentctl validate` warns about any tool parameter without a description, or with a blank one. It does not fail validation yet. Parameters LangChain fills in at runtime, such as `config: RunnableConfig` or `InjectedToolArg` annotations, are never sent to the model and are not checked.
 
 <Tip>
   Write clear descriptions and typed parameters. The LLM reads both when deciding how to call a tool. Vague descriptions lead to wrong or missed tool calls.
diff --git a/src/agent_engine/core/tool_metadata.py b/src/agent_engine/core/tool_metadata.py
index d5415e47..94665b3b 100644
--- a/src/agent_engine/core/tool_metadata.py
+++ b/src/agent_engine/core/tool_metadata.py
@@ -17,6 +17,10 @@ def get_customer(customer_code: Annotated[str, "The unique customer code."]) ->
 (``ast``), never executed. Anything the scan cannot positively identify — a
 missing or unparseable file, a function it cannot find — is skipped and left
 to the existing plugin checks and runtime loaders to report.
+
+Parameters LangChain fills in at runtime are left out of the schema sent to the
+model, so they are not checked. Being static, the scan recognizes them by the
+names they are written with (see ``_is_runtime_injected``).
 """
 
 from __future__ import annotations
@@ -63,6 +67,8 @@ def _check_tool(path: Path, tool_id: str) -> list[str]:
         return []
     warnings: list[str] = []
     for arg in (*fn.args.posonlyargs, *fn.args.args, *fn.args.kwonlyargs):
+        if _is_runtime_injected(arg):
+            continue
         description = _annotated_description(arg.annotation)
         if description is None:
             warnings.append(f'tool "{tool_id}": parameter "{arg.arg}" is missing a description')
@@ -71,6 +77,54 @@ def _check_tool(path: Path, tool_id: str) -> list[str]:
     return warnings
 
 
+# LangChain drops these parameter names from every tool schema (``FILTERED_ARGS``).
+_RUNTIME_ARG_NAMES = frozenset({"callbacks", "run_manager"})
+
+
+def _is_runtime_injected(arg: ast.arg) -> bool:
+    """True for a parameter LangChain fills in itself and never shows the model.
+
+    Mirrors ``langchain_core`` tool-schema filtering:
+
+    - ``callbacks`` / ``run_manager``, by name;
+    - ``config: RunnableConfig`` — only the bare type; ``RunnableConfig | None``
+      is still shown to the model;
+    - ``ToolRuntime`` or ``ToolRuntime[...]``;
+    - ``Annotated[T, Injected...]``: ``InjectedToolArg`` and its subclasses
+      (``InjectedToolCallId``, LangGraph's ``InjectedState`` / ``InjectedStore``),
+      bare or called.
+    """
+    if arg.arg in _RUNTIME_ARG_NAMES:
+        return True
+    annotation = arg.annotation
+    if annotation is None:
+        return False
+    if _type_name(annotation) in {"RunnableConfig", "ToolRuntime"}:
+        return True
+    if isinstance(annotation, ast.Subscript):
+        if _type_name(annotation.value) == "ToolRuntime":
+            return True
+        if _is_annotated(annotation.value) and isinstance(annotation.slice, ast.Tuple):
+            return any(_is_injected_marker(item) for item in annotation.slice.elts[1:])
+    return False
+
+
+def _is_injected_marker(node: ast.expr) -> bool:
+    """``InjectedState``, ``InjectedState("messages")``, ``tools.InjectedToolArg()``…"""
+    target = node.func if isinstance(node, ast.Call) else node
+    name = _type_name(target)
+    return name is not None and name.startswith("Injected")
+
+
+def _type_name(node: ast.expr) -> str | None:
+    """The bare name of ``X`` or ``module.X``; None for anything else."""
+    if isinstance(node, ast.Name):
+        return node.id
+    if isinstance(node, ast.Attribute):
+        return node.attr
+    return None
+
+
 def _annotated_description(annotation: ast.expr | None) -> str | None:
     """The first string in ``Annotated[T, ...]`` metadata, or None if there is none."""
     if not isinstance(annotation, ast.Subscript) or not _is_annotated(annotation.value):
diff --git a/tests/cli/test_validate_command.py b/tests/cli/test_validate_command.py
index 9f393b1a..e021648e 100644
--- a/tests/cli/test_validate_command.py
+++ b/tests/cli/test_validate_command.py
@@ -319,6 +319,22 @@ def test_described_tool_parameter_has_no_warnings(tmp_path: Path) -> None:
     assert "warnings:" not in CliRunner().invoke(cli, ["validate", spec]).output
 
 
+def test_runtime_injected_parameter_is_not_reported(tmp_path: Path) -> None:
+    """LangChain fills ``config`` in at runtime; the model never sees it."""
+    spec = _spec_with_tool(
+        tmp_path,
+        "from typing import Annotated\n"
+        "from langchain_core.runnables import RunnableConfig\n\n"
+        "def lookup(query: Annotated[str, 'What to look up.'], config: RunnableConfig) -> str:\n"
+        "    return query\n",
+    )
+
+    result = validate_spec(spec)
+
+    assert result.ok, result.errors
+    assert result.warnings == []
+
+
 def test_warnings_are_shown_alongside_errors(tmp_path: Path) -> None:
     spec = _spec_with_tool(tmp_path, "def lookup(query: str) -> str:\n    return query\n")
     # Break the spec so validation fails for an unrelated reason.
diff --git a/tests/core/test_tool_metadata.py b/tests/core/test_tool_metadata.py
index 48339803..79855f88 100644
--- a/tests/core/test_tool_metadata.py
+++ b/tests/core/test_tool_metadata.py
@@ -171,6 +171,101 @@ def test_star_args_and_kwargs_are_ignored(tmp_path: Path) -> None:
     assert warnings == []
 
 
+# -- runtime-injected parameters: LangChain fills these in; the model never sees them --
+
+
+def test_runnable_config_parameter_is_not_reported(tmp_path: Path) -> None:
+    """The case from the #156 review: the model's schema only exposes ``query``."""
+    warnings = scan_one(
+        tmp_path,
+        """\
+        from typing import Annotated
+        from langchain_core.runnables import RunnableConfig
+
+        def get_customer(query: Annotated[str, "What to look up."], config: RunnableConfig) -> str:
+            return query
+        """,
+    )
+
+    assert warnings == []
+
+
+def test_optional_runnable_config_is_still_reported(tmp_path: Path) -> None:
+    """LangChain only hides a bare ``RunnableConfig``; an optional one reaches the model."""
+    warnings = scan_one(
+        tmp_path,
+        """\
+        from langchain_core.runnables import RunnableConfig
+
+        def get_customer(config: RunnableConfig | None = None) -> str:
+            return ""
+        """,
+    )
+
+    assert warnings == ['tool "get_customer": parameter "config" is missing a description']
+
+
+def test_callback_parameters_are_not_reported(tmp_path: Path) -> None:
+    warnings = scan_one(
+        tmp_path,
+        "def get_customer(callbacks=None, run_manager=None) -> str:\n    return ''\n",
+    )
+
+    assert warnings == []
+
+
+def test_tool_runtime_is_not_reported(tmp_path: Path) -> None:
+    warnings = scan_one(
+        tmp_path,
+        """\
+        from langchain.tools import ToolRuntime
+
+        def get_customer(runtime: ToolRuntime, other: ToolRuntime[dict, dict]) -> str:
+            return ""
+        """,
+    )
+
+    assert warnings == []
+
+
+def test_injected_annotations_are_not_reported(tmp_path: Path) -> None:
+    warnings = scan_one(
+        tmp_path,
+        """\
+        from typing import Annotated
+        from langchain_core import tools
+        from langchain_core.tools import InjectedToolArg, InjectedToolCallId
+        from langgraph.prebuilt import InjectedState, InjectedStore
+
+        def get_customer(
+            call_id: Annotated[str, InjectedToolCallId],
+            arg: Annotated[str, InjectedToolArg()],
+            qualified: Annotated[str, tools.InjectedToolArg],
+            state: Annotated[dict, InjectedState],
+            messages: Annotated[list, InjectedState("messages")],
+            store: Annotated[object, InjectedStore()],
+        ) -> str:
+            return ""
+        """,
+    )
+
+    assert warnings == []
+
+
+def test_injected_parameters_do_not_hide_undescribed_model_parameters(tmp_path: Path) -> None:
+    warnings = scan_one(
+        tmp_path,
+        """\
+        from langchain_core.runnables import RunnableConfig
+
+        def get_customer(query: str, config: RunnableConfig) -> str:
+            return query
+        """,
+    )
+
+    assert warnings == ['tool "get_customer": parameter "query" is missing a description']
+
+
 def test_async_tool_is_scanned(tmp_path: Path) -> None:
     warnings = scan_one(
         tmp_path,