anthropic_llm - bugs/missing support

[marcfong]

Describe the bug
A clear and concise description of what the bug is.
src/google/adk/models/anthropic_llm.py

Using Claude models with ADK, i've observed the following issues:

1. No streaming support
2. No support for embedded attachments in message part
3. No support for function responses that have complex python ARRAY DICTs

To Reproduce
Please share a minimal code and data to reproduce your problem.
Steps to reproduce the behavior:
2 - Add image parts to a user message and send to the ADK agent, it will fail with error.

3 - Add a python FunctionTool which returns a DICT with attributes which are arrays or complex objects. Even though ADK appears to work, Claude does not receive the response/result payload from after the function call happens. THis is different from how Gemini behaves and forces refactoring of FunctionTool to only return simple text responses.

Expected behavior
Claude on ADK supports the same capabilities as Gemini in ADK.

Screenshots
If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: [e.g. macOS, Linux, Windows]
• Python version(python -V):
• ADK version(pip show google-adk):

Model Information:

• Are you using LiteLLM: Anthropic LLM
• Which model is being used(e.g. gemini-2.5-pro) : Clause 4.5 Sonnet

Additional context
Add any other context about the problem here.

[alfredolozano]

+1!

[gplhegde]

Seeing the same issue. Even if the response of the function tool is very simple dict, it is not passed to the LLM in the tool_result content. This is odd. Below is the code and the LLM call log

Sample agent

import random

from google.adk.agents import LlmAgent
from google.adk.tools import FunctionTool
from google.genai.types import GenerateContentConfig

from timblai.registry import registry

# MODEL = "gemini-2.5-flash"
MODEL = "claude-haiku-4-5@20251001"


def get_topic() -> str:
    """
    Get a random joke topic.

    Returns:
        str: A joke topic.
    """
    topics = [
        "technology",
        "animals",
        "school",
        "food",
        "sports",
        "work",
        "relationships",
        "travel",
        "music",
    ]
    return {"topic": random.choice(topics)}


root_agent = LlmAgent(
    name="joker_agent",
    description="Tells jokes based on user input.",
    instruction="You are a joker agent. You should call `get_topic` tool to get the topic and generate joke every time a user asks for a joke and end your turn.",
    model=MODEL,
    tools=[
        FunctionTool(get_topic),
    ],
    generate_content_config=GenerateContentConfig(max_output_tokens=400),
)

Partial call log 👇

Claude response: {
  "id": "msg_vrtx_01VomiarqErJxjghnHqEq3TY",
  "content": [
    {
      "id": "toolu_vrtx_014rCL3NTSNqRhqp1zwRCBQu",
      "input": {},
      "name": "get_topic",
      "type": "tool_use"
    }
  ],
  "model": "claude-haiku-4-5-20251001",
  "role": "assistant",
  "stop_reason": "tool_use",
  "type": "message",
  "usage": {
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "input_tokens": 635,
    "output_tokens": 37
  }
}
2025-10-24 09:19:22,918 - DEBUG - adk_web_server.py:1399 - Generated event in agent run streaming: {"content":{"parts":[{"functionCall":{"id":"toolu_vrtx_014rCL3NTSNqRhqp1zwRCBQu","args":{},"name":"get_topic"}}],"role":"model"},"usageMetadata":{"candidatesTokenCount":37,"promptTokenCount":635,"totalTokenCount":672},"invocationId":"e-8113ae6f-05e6-4147-bdb8-39a0db0ce767","author":"joker_agent","actions":{"stateDelta":{},"artifactDelta":{},"requestedAuthConfigs":{},"requestedToolConfirmations":{}},"longRunningToolIds":[],"id":"b28472ad-7e6b-4c9a-8d48-cc81df1e4cce","timestamp":1761268760.450154}
2025-10-24 09:19:22,919 - DEBUG - adk_web_server.py:1399 - Generated event in agent run streaming: {"content":{"parts":[{"functionResponse":{"id":"toolu_vrtx_014rCL3NTSNqRhqp1zwRCBQu","name":"get_topic","response":{"topic":"travel"}}}],"role":"user"},"invocationId":"e-8113ae6f-05e6-4147-bdb8-39a0db0ce767","author":"joker_agent","actions":{"stateDelta":{},"artifactDelta":{},"requestedAuthConfigs":{},"requestedToolConfirmations":{}},"id":"fed6a43d-9c14-49a4-9bd5-ddfdec487486","timestamp":1761268762.91918}
2025-10-24 09:19:22,926 - DEBUG - _base_client.py:490 - Request options: {'method': 'post', 'url': '/projects/fastq2c/locations/us-east5/publishers/anthropic/models/claude-haiku-4-5@20251001:rawPredict', 'timeout': Timeout(connect=5.0, read=600, write=600, pool=600), 'files': None, 'idempotency_key': 'stainless-python-retry-260880a2-8bfc-4762-abcb-7e4be17fc5e8', 'json_data': {'max_tokens': 8192, 'messages': [{'role': 'user', 'content': [{'text': 'tell me a good joke', 'type': 'text'}]}, {'role': 'assistant', 'content': [{'id': 'toolu_vrtx_014rCL3NTSNqRhqp1zwRCBQu', 'name': 'get_topic', 'input': {}, 'type': 'tool_use'}]}, {'role': 'user', 'content': [{'tool_use_id': 'toolu_vrtx_014rCL3NTSNqRhqp1zwRCBQu', 'type': 'tool_result', 'content': '', 'is_error': False}]}], 'system': 'You are a joker agent. You should call `get_topic` tool to get the topic and generate joke every time a user asks for a joke and end your turn.\n\nYou are an agent. Your internal name is "joker_agent".\n\n The description about you is "Tells jokes based on user input."', 'tool_choice': {'type': 'auto'}, 'tools': [{'name': 'get_topic', 'description': '\n    Get a random joke topic.\n\n    Returns:\n        str: A joke topic.\n    ', 'input_schema': {'type': 'object', 'properties': {}}}], 'anthropic_version': 'vertex-2023-10-16'}}

See the tool_result content is empty in the final part of the json_data.

Even the config parameters like max_tokens is not getting passed.

Possible short term workaround - json.dumps the dict from the tool response.

I am using adk, version 1.16.0

[shankyram1912]

• 1, it is very important for many of our customers to have 1st class support for Anthropic models with ADK

[Happy-ryan]

Hi team. I've addressed the image attachment issue (point 2) in #3286 🙏

[surajksharma07]

@marcfong The function response bug is ADK converting dicts with str() instead of JSON.
Quick workaround: import json and return json.dumps(your_dict) instead of the dict directly - Claude will receive it properly then.
Streaming support is coming (tracked in TODO), and PR #3286 fixes images.
We're looking more into it for all three.
Thanks for the detailed report!