[BUG] Bedrock/Claude model incorrectly serializes nested object parameters as JSON string causing Pydantic validation errors

[skarltjr]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.19.0

Python Version

3.11.9

Operating System

macos 14.6.1

Installation Method

other

Steps to Reproduce

Bedrock/Claude model sends nested object parameters as JSON string instead of dict

Description

When using strands with MCP tools that have nested object parameters (e.g., parameters: dict), the Bedrock/Claude model incorrectly serializes the parameter value as a JSON string instead of passing it as a native Python dictionary.

This issue does not occur with the Gemini model, which correctly passes nested objects as dictionaries.

Environment

• strands version: 1.19.0
• fastmcp version: 2.13.2 / 2.13.3
• Python version: 3.11 / 3.12
• Model (problematic): us.anthropic.claude-sonnet-4-20250514-v1:0 (Bedrock) & other claude model versions
  • "arn:aws:bedrock:us-east-2::inference-profile/global.anthropic.claude-opus-4-5-20251101-v1:0"
  • ... other claude models
• Model (working): gemini-2.5-flash (Google AI)

Tool Schema (from MCP Server)

{
  "name": "ecs_troubleshooting_tool",
  "inputSchema": {
    "properties": {
      "action": {
        "type": "string",
        "enum": ["fetch_service_events", "fetch_task_failures", "..."]
      },
      "parameters": {
        "anyOf": [
          { "additionalProperties": true, "type": "object" },
          { "type": "null" }
        ],
        "default": null
      }
    },
    "type": "object"
  }
}

Expected Behavior

The parameters argument should be passed as a Python dictionary:

ecs_troubleshooting_tool(
    action="fetch_service_events",
    parameters={
        "ecs_cluster_name": "my-cluster",
        "ecs_service_name": "my-service",
        "time_window": 3600
    }
)

Actual Behavior (Bedrock/Claude)

The parameters value is serialized as a JSON string:

ecs_troubleshooting_tool(
    action="fetch_service_events",
    parameters='{"ecs_cluster_name": "my-cluster", "ecs_service_name": "my-service", "time_window": 3600}'
)

This causes a Pydantic validation error:

ValidationError: 1 validation error for call[mcp_ecs_troubleshooting_tool]
parameters
  Input should be a valid dictionary [type=dict_type, input_value='{"ecs_cluster_name": "in...dol-test-vcs-use2-dev"}', input_type=str]

Tool #5: ecs_troubleshooting_tool
[12/04/25 13:18:48] Error validating tool 'ecs_troubleshooting_tool': 1 validation error for call[mcp_ecs_troubleshooting_tool]                                                                            
                    parameters                                                                                                                                                                             
                      Input should be a valid dictionary [type=dict_type, input_value='{"ecs_cluster_name": "in...", "time_window": 7200}', input_type=str]                                                
                        For further information visit https://errors.pydantic.dev/2.12/v/dict_type                                                                                                         
                    ╭───────────────────────────────────────────────────────────────────────── Traceback (most recent call last) ─────────────────────────────────────────────────────────────────────────╮
                    │ /Users/a18144/.cache/uv/archive-v0/JtV7en25lIaDnPI8k48sK/lib/python3.11/site-packages/fastmcp/tools/tool_manager.py:160 in call_tool                                                │
                    │                                                                                                                                                                                     │
                    │ /Users/a18144/.cache/uv/archive-v0/JtV7en25lIaDnPI8k48sK/lib/python3.11/site-packages/fastmcp/tools/tool.py:341 in run                                                              │
                    │                                                                                                                                                                                     │
                    │ /Users/a18144/.cache/uv/archive-v0/JtV7en25lIaDnPI8k48sK/lib/python3.11/site-packages/pydantic/type_adapter.py:441 in validate_python                                               │
                    ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
                    ValidationError: 1 validation error for call[mcp_ecs_troubleshooting_tool]                                                                                                             
                    parameters                                                                                                                                                                             
                      Input should be a valid dictionary [type=dict_type, input_value='{"ecs_cluster_name": "in...", "time_window": 7200}', input_type=str]   

Steps to Reproduce

1. Create an MCP tool with a nested object parameter (e.g., parameters: dict | None)
2. Use strands with BedrockModel (Claude)
3. Have the agent call the tool with nested parameters
4. Observe the validation error

Workaround

Switching to Gemini model resolves the issue, as Gemini correctly passes nested objects as dictionaries.

What I've Tried

• Prompt engineering to instruct the model to "pass parameters as an object, not a string" — did not work

Possible Root Causes

1. BedrockModel response parsing: The tool call arguments from Bedrock Converse API may be returned as a JSON-encoded string, and strands doesn't deserialize nested objects.

2. Schema interpretation: Claude may interpret type: "object" differently and choose to stringify nested structures.

3. strands tool invocation: The library may not handle the case where nested parameters need to be parsed from JSON strings.

Potentially Related Files

• strands/models/bedrock.py — BedrockModel implementation
• strands/event_loop/ — Tool execution logic

Expected Behavior

..

Actual Behavior

..

Additional Context

No response

Possible Solution

No response

Related Issues

No response

[agent-of-mkmeral]

Hi — we traced through this and confirmed it's a real bug on strands-agents==1.19.0, but it was already fixed in v1.20.0 by PR #1301.

Fix: Upgrade to strands-agents>=1.20.0.

---

🐛 Root cause (what was broken in 1.19.0)

In _normalize_property() (v1.19.0), the code unconditionally adds type: "string" to every property that doesn't already have a top-level type:

# v1.19.0 - BROKEN
normalized_prop.setdefault("type", "string")  # Always runs

For MCP tool properties with anyOf: [{type: "object"}, {type: "null"}] (which have no top-level type field), this creates a contradictory schema sent to Claude:

{
  "anyOf": [{"additionalProperties": true, "type": "object"}, {"type": "null"}],
  "type": "string",
  "default": null
}

Claude sees both type: "string" AND anyOf: [{type: "object"}, {type: "null"}] — and sometimes resolves the contradiction by returning the nested object as a JSON string.

✅ The fix (v1.20.0+, PR #1301)

# v1.20.0+ - FIXED
has_composition = any(kw in normalized_prop for kw in _COMPOSITION_KEYWORDS)
if not has_composition:
    normalized_prop.setdefault("type", "string")  # Only adds if no anyOf/oneOf/allOf/not

Now if a property has anyOf, oneOf, allOf, or not, the type: "string" default is not injected. Claude sees a clean schema and returns proper dicts.

🧪 Verification (18/18 tests pass on latest)

Environment: strands-agents 1.37.0, fastmcp 3.2.4, Python 3.13, Claude Sonnet on Bedrock

We ran 18 test cases (exact scenario from your report, nested dicts, multi-step, null params, rapid-fire multi-tool, etc.) — all passed with proper dict objects returned.

Version	Schema sent to Claude	Bug?
1.19.0	{anyOf: [...], type: "string", ...}	✅ Yes — contradictory
1.20.0+	{anyOf: [...], ...} (no spurious type)	❌ Fixed

---

TL;DR: This is fixed. Upgrade to strands-agents>=1.20.0 (latest is 1.37.0) and the issue should be resolved.

---

🤖 AI agent response. Strands Agents. Feedback welcome!

[mehtarac]

Closing this issue as the fix has been merged: #1301