Azure AI Foundry Agent works in Playground but fails in Microsoft 365 Copilot when using Knowledge Base

Welcome to Microsoft Q&A

Hello Paul Satterlee,

You have run into a documented architectural limitation regarding how Microsoft 365 Copilot renders payloads from Azure AI Foundry agents.The reason your trace shows Chat/LLM = Success but the Copilot UI throws a "Sorry, something went wrong" error comes down to citations and streaming.

1. Copilot enforces a stricter rendering pipeline than the Foundry playground; an agent can complete tool calls and LLM steps in Foundry traces yet still be rejected by Copilot at render time if the final payload uses unsupported features.
2. Foundry IQ knowledge sources are separate from Copilot’s native knowledge sources; retrieval results and citation metadata produced by Foundry may not map to Copilot’s expected rendering model.
3. Community reports and technical notes also list token/system‑prompt differences and other platform limitations that can change behavior between environments.

When your Knowledge Base retrieval tool executes, the LLM processes the injected context and typically formats its final response with grounding citations (footnotes) and attempts to stream the output. However, according to the official Publish agents to Microsoft 365 Copilot documentation, published agents do not support streaming responses or citations within the Microsoft 365 Copilot canvas.

When the Copilot rendering engine receives the Foundry-formatted payload containing citation metadata or a streaming flag, it fails to parse the structure. Instead of failing gracefully, it aborts the UI render entirely, resulting in the generic error message you are seeing.

To get your Knowledge Base agent working in Copilot, you must strip out the unsupported metadata before the payload hits the Copilot UI.

1. Simplify the final response — return plain, non‑streamed text with no embedded citation blocks or HTML. If Copilot shows the answer, the issue is rendering.
2. Disable streaming and citation output in the agent’s response template or post‑processing step. Copilot does not support streaming or some citation formats for published agents.
3. Strip tool metadata (source URLs, JSON blobs, attachments) from the final message; include only the human‑readable answer.
4. Test retrieval settings — set retrieval reasoning to minimal/low and return only the synthesized answer (not raw document extracts).
5. Reduce token footprint and avoid hidden system prompt expansions that may push payloads over Copilot limits.

By forcing the agent to return a single, plain-text block without citation tags, Microsoft 365 Copilot will be able to render the response successfully.

😊 If my answer helped you resolve your issue, please consider marking it as the correct answer. This helps others in the community find solutions more easily. Thanks!