Anthropic function limit fallback can return empty final response #5769

Summary

When the function invocation loop reaches max_function_calls or max_iterations, the core loop correctly stops additional tool calls and makes a final no-tool model call. With the Anthropic provider, that final fallback call can return stop_reason=end_turn with an empty content array, producing a blank final response.

Why this matters

The safety limit itself works: tool calls are capped. The problem is the user-visible final turn. A production agent can hit a function-call/iteration limit and return no text instead of explaining that the limit was reached.

Observed behavior

Using agent-framework-anthropic with Anthropic Opus 4.7:

• Tool function returns call again until called 3 times.
• Client config sets either:
  • client.function_invocation_configuration["max_function_calls"] = 1, or
  • client.function_invocation_configuration["max_iterations"] = 1
• SDK executes exactly one tool call, then performs the final fallback call with tool_choice={"type":"none"}.
• Raw Anthropic final response is:
  • stop_reason="end_turn"
  • content=[]

The resulting Agent.run(...) response has no text.

Relevant SDK behavior

The core loop appears to intend a final no-tool answer after limit exhaustion:

• Non-streaming path sets mutable_options["tool_choice"] = "none" and calls the underlying chat client again after loop exhaustion.
• Streaming path does the same with a final inner stream.
• Core tests mock this final call as producing fallback text.

That assumption does not always hold for Anthropic when the last tool result says the model should continue.

Minimal shape of the final Anthropic request

The final request sent by the SDK looked like this conceptually:

{
  "tool_choice": {"type": "none"},
  "tools": [{"type": "custom", "name": "continue_until_three", "input_schema": {}}],
  "messages": [
    {"role": "user", "content": [{"type": "text", "text": "Start the loop."}]},
    {"role": "assistant", "content": [{"type": "tool_use", "id": "toolu_...", "name": "continue_until_three", "input": {"marker": "1"}}]},
    {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "toolu_...", "content": [{"type": "text", "text": "call again"}], "is_error": false}]}
  ]
}

Anthropic returned content=[].

Direct Anthropic comparison showed:

• Same messages + tool_choice=none -> empty content.
• Same messages with no tools -> empty content.
• Same messages plus explicit user instruction like Tool call limit reached. Do not call more tools. Explain what happened in one sentence. -> text response.

Expected behavior

When the function invocation loop stops because a configured limit is reached, the SDK should not return an empty final assistant message. Possible fixes:

1. Append an explicit final instruction before the no-tool fallback call, e.g. explain that the tool/iteration limit was reached and produce a final answer without further tool use.
2. If the fallback call still returns empty content, synthesize a deterministic limit-reached assistant message.
3. Expose enough metadata to callers so they can reliably detect limit_reached and apply their own deterministic fallback.

Environment

• Python SDK from microsoft/agent-framework local clone
• agent-framework-anthropic dependency pin locally loosened to allow anthropic>=0.97.0,<1
• anthropic==0.101.0
• Model: claude-opus-4-7
• Thinking: {"type":"adaptive","display":"omitted"}
• Output config: {"effort":"low"}

Local proof

This was found while evaluating Microsoft Agent Framework adoption in a production Azure Functions assistant runtime. We are keeping a local guard to avoid returning blank user responses when a tool/iteration limit is reached.