Claude models fail silently in Foundry Agent Service MCP pipeline — tool results not propagated back to model

Hello @Vivek Kumar (Admin)

Thank you for providing such a detailed description of the issue and the troubleshooting you've already performed.

From your findings, I understand that the MCP integration itself appears to be functioning correctly. The OAuth authentication flow completes successfully, the mcp_list_tools operation returns the available Dynamics 365 Finance & Operations tools, and individual MCP tool calls such as data_find_entity_type and data_get_entity_metadata execute successfully and return the expected data. However, when using Claude Sonnet 4.5/4.6 in Azure AI Foundry Agent Service, the returned tool results are not propagated back to the model for the next reasoning step. Instead, the agent run ends with a server_error, and the trace shows events: [] at the invoke_agent span level. In contrast, GPT-5 completes the same multi-step MCP workflow successfully using the identical MCP configuration.

Based on the information you've shared, this suggests that the MCP connectivity, authentication, and Dynamics 365 ERP MCP server are functioning correctly. The behavior appears to occur during the agent runtime after the tool execution completes, where the tool results are expected to be passed back to the model for continued reasoning.

Since GPT-5 works successfully with the same MCP server, OAuth configuration, prompts, and tool definitions, this makes it less likely that the issue is related to your MCP server implementation or authentication configuration. Instead, it points toward a model-specific behavior or a limitation in how the Azure AI Foundry Agent Service currently handles tool result propagation for the Claude models.

Additionally, Microsoft documentation recommends Claude Sonnet models for the Dynamics 365 ERP MCP server scenario, so the workflow you're implementing is a supported scenario. Based on the available documentation, however, there is no known public guidance describing an issue where Claude models fail to consume MCP tool results after successful tool execution.

Before concluding that this is a platform issue, we recommend verifying the following:

• Confirm that the Claude model deployment is healthy and available in your Azure AI Foundry project and region.
• Review the Agent Service traces and monitoring data to determine whether any additional runtime errors are recorded after the MCP tool completes.
• Verify that Application Insights (if configured) and Agent Service diagnostics are capturing the complete execution flow.
• Compare the trace outputs between a successful GPT-5 execution and a failing Claude execution to identify where the execution path diverges after the MCP tool returns its response.
• Confirm that the same MCP tool definitions, prompts, authentication, and conversation state are being used across both GPT-5 and Claude models, since you've observed GPT-5 completing the workflow successfully.

Please do let me know if you have further queries.

Thank you!

Great troubleshooting and clear problem analysis! The way you isolated the issue by comparing Claude and GPT-5 with the same MCP configuration is really impressive. The evidence points toward the tool-result handoff layer, which is a valuable observation.

I faced a similar type of MCP integration issue before, and based on the symptoms it looks less like an F&O MCP server problem and more like a model/tool-calling compatibility issue in the agent runtime layer.

Since:

• OAuth flow is successful
• "mcp_list_tools" works
• MCP tools execute successfully and return data
• GPT-5 completes the same tool chain with the same configuration

the failure point seems to be the handoff of the MCP tool response back into the Claude reasoning cycle.

A few things I would check:

1. Verify tool response schema compatibility
   • Compare the MCP tool output format with what Claude expects.
   • Some models can be more strict about structured tool results (JSON schema, content blocks, empty fields, etc.).
2. Check Foundry Agent Service runtime/model version
   • If GPT-5 works and Claude fails with identical tools, there may be a runtime bug or limitation in the current Claude model integration.
3. Try simplifying the MCP response
   • Temporarily return a smaller payload from F&O tools.
   • Remove optional fields/large metadata and test if Claude continues.
4. Review the trace at the tool-result boundary
   • The important clue is "events: []".
   • The tool executes, but the result is not being attached to the agent turn, which suggests the failure occurs after tool execution and before model continuation.

Because Microsoft documentation lists Claude Sonnet for this MCP scenario, I would raise this as a Foundry Agent Service issue with the trace IDs. The evidence already points to a platform-side problem rather than an authentication or MCP configuration issue.

A workaround for now could be using GPT-5 for this agent flow until the Claude tool-result propagation issue is resolved.

Please do let me know if you have further queries.

Thank you!

Hi @Vivek Kumar (Admin)

Did you get any chance to review the above response. Do let me know if you have any further queries.

Thank you!

Hi @Vivek Kumar (Admin)

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thank you!