fix(agent): stop writing the tool schemas into the conversation

tool_handling() appended tools_list_dictionary to short_memory under
role=self.agent_name. The same schemas already reach the provider through the
API's tools parameter (LLMManager.build passes tools_list_dictionary to
LiteLLM), so every request carried them twice, attributed to an assistant turn
the assistant never produced, and stored as a list where every other entry is a
string.

Closes kyegomez#1989

Author: ayaangazali

swarms/structs/agent.py

@@ -897,7 +897,6 @@ def tool_handling(self):
         2. Initializes tools_list_dictionary if None
         3. Tracks existing tool names to prevent duplicates
         4. Adds new tools that don't already exist
-        5. Adds tools to conversation memory for LLM context
 
         **Tool Schema Format:**
         Tools are converted to OpenAI function calling format:
@@ -923,16 +922,16 @@ def tool_handling(self):
             None: Uses self.tools and self.tools_list_dictionary from instance.
 
         Returns:
-            None: Modifies self.tools_list_dictionary and self.short_memory.
+            None: Modifies self.tools_list_dictionary.
 
         Note:
             - This method is called automatically during agent initialization if tools are provided
-            - Tools are added to conversation memory so the LLM knows what tools are available
+            - Schemas reach the model through the API's `tools` parameter, not the conversation
             - The method preserves existing tools in tools_list_dictionary (e.g., handoff tools)
             - Tool names are case-sensitive for duplicate detection
 
         Raises:
-            Exception: If tool conversion fails or tools cannot be added to memory.
+            Exception: If tool conversion fails.
 
         Examples:
             >>> def my_tool(query: str) -> str:
@@ -967,11 +966,6 @@ def tool_handling(self):
                 self.tools_list_dictionary.append(tool)
                 existing_tool_names.add(tool_name)
 
-        self.short_memory.add(
-            role=self.agent_name,
-            content=self.tools_list_dictionary,
-        )
-
     def short_memory_init(self):
         # Compactly assemble initial prompt as a string with available fields
         prompt = self.system_prompt

tests/structs/test_agent.py

@@ -862,6 +862,32 @@ def test_default_is_per_instance(self):
         )
 
 
+class TestToolSchemasStayOutOfTheConversation:
+    """Schemas reach the model as the API's `tools` parameter, not as a turn."""
+
+    def test_no_conversation_turn_carries_the_tool_schemas(self):
+        # dynamic_tools=False is the branch that reaches tool_handling().
+        agent = _patched_agent(
+            "SchemaAgent",
+            model_name="gpt-5.4",
+            dynamic_tools=False,
+            tools=[_math_tool],
+        )
+
+        assert any(
+            schema["function"]["name"] == "_math_tool"
+            for schema in agent.tools_list_dictionary
+        )
+        assert [
+            m
+            for m in agent.short_memory.conversation_history
+            if not isinstance(m["content"], str)
+        ] == []
+        assert [
+            m["role"] for m in agent.short_memory.conversation_history
+        ] == ["System"]
+
+
 class TestAutonomousAgentLoop:
     """The ``max_loops="auto"`` loop lives in ``AutonomousAgentLoop``, not on
     ``Agent``. These pin the seam: the loop is wired up, it reads and writes