Regression in Copilot Studio output rendering for long structured responses (started this week)

This behavior aligns with a platform‑side change affecting how long, complex outputs are rendered rather than how they are generated.

For long, structured responses, the main controllable factor is output length. Longer outputs increase the chance of truncation, collapsing, or malformed rendering in the chat UI, even when the full text is present in traces.

To mitigate the regression while the platform behavior is in this state:

1. Constrain the length of each response
   • Explicitly instruct the agent to keep answers within a fixed size, for example:
     • “Summarize in 50–150 words per section.”
     • “Limit each section to 5 bullet points.”
   • This reduces both response size and rendering complexity, which improves how the UI displays the content.
2. Reduce structural complexity in the output
   • Avoid deeply nested or very wide structures (for example, large tables with many columns).
   • Prefer shorter, simpler bullet lists and headings instead of large tables when possible.
   • When JSON or other structured formats are used, simplify the structure and minimize the number of keys so the rendered text is more compact.
3. Only include information that is actually needed
   • Remove optional or rarely used sections from the generated output.
   • Avoid asking the model to produce explanations or rationales that are not required for the governance review, as these add length and increase the chance of UI issues.
4. Consider a two‑step pattern for very long reviews
   • First response: a concise, high‑level assessment (for example, key risks, pass/fail, major issues) with strict length limits.
   • Follow‑up responses: on demand, ask the agent to “provide more details” for a specific section or risk area. This keeps each individual message shorter and more reliably rendered.
5. Use summarization to manage verbosity during review
   • If the agent produces a long answer that is hard to read in the UI, immediately ask it to “summarize the response” or “summarize in N bullet points.”
   • This pattern is explicitly supported and helps keep responses concise and readable in live conversations.

These steps do not change the underlying model behavior but adapt the prompts and output format to current platform constraints, improving readability and reducing the likelihood of cut‑off or collapsed rendering for long, structured document‑review responses.

References:

• Prompts performance and execution
• Responsible AI FAQ for Copilot in Customer Service