Fix JSONDecodeError when tool_call.arguments is empty string in AssistantAgent._execute_tool_call

[linger118927]

Description

When using Claude models (or other LLM providers) with tool calling, the tool_call.arguments field may be an empty string "" instead of "{}" for tools that don't require parameters.

This causes a JSONDecodeError[ in ](file://core/mt_llm_client.py#90#24)AssistantAgent._execute_tool_call method, even though the tool call is valid and should be executed with empty arguments.

Current Behavior

# In _assistant_agent.py line 1547
arguments = json.loads(tool_call.arguments)  # Raises JSONDecodeError for ""

When tool_call.arguments = "", this raises:

Why are these changes needed?

Related issue number

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. See https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

[nidhishgajjar]

Orb Code Review (powered by GLM 5.1 on Orb Cloud)

Summary

This PR fixes a JSONDecodeError that occurs when tool_call.arguments is an empty string. Some LLM providers return an empty string for tool calls with no parameters, and json.loads('') raises a JSONDecodeError. The fix checks for empty/whitespace-only arguments and substitutes an empty dict instead.

Architecture

The _execute_tool_call static method in AssistantAgent is responsible for parsing tool call arguments from JSON and dispatching the call. FunctionCall.arguments is typed as str (a JSON string), so callers must handle the deserialization. The fix is placed right before the json.loads call, which is the correct location.

Issues

Warning — Redundant empty check

File: python/packages/autogen-agentchat/src/autogen_agentchat/agents/_assistant_agent.py:1547-1548
Severity: suggestion

The condition not tool_call.arguments or tool_call.arguments.strip() == "" has some redundancy. not tool_call.arguments already covers None, "", and other falsy strings. tool_call.arguments.strip() == "" additionally catches whitespace-only strings like " ". You could simplify to:

if not tool_call.arguments or not tool_call.arguments.strip():

or even more concisely:

if not tool_call.arguments or tool_call.arguments.isspace():

This is minor and the current form works correctly — just noting for readability.

Note — FunctionCall.arguments typing

Severity: suggestion

FunctionCall.arguments is typed as str in autogen_core/_types.py (a dataclass with no default). It might be worth considering whether the type should allow Optional[str] or if the constructors in _openai_client.py and _azure_ai_client.py should default to '{}' instead of ''. That would be a broader fix at the source, but this PR's approach of handling it at the consumption site is perfectly reasonable and more defensive.

Cross-file impact

None directly. The fix is contained within _execute_tool_call. Related code paths that construct FunctionCall objects with empty arguments exist in openai/_openai_client.py and azure/_azure_ai_client.py but those don't need changes for this fix.

Assessment

Approve ✅ — Clean, focused fix for a real bug. Empty tool call arguments from certain LLM providers would crash the agent, and this handles it gracefully. The approach is defensive and correct.