Forum Discussion
Fabric Data Agent performance is not the same via Copilot Studio agent
Hi kevinvu99
In my opinion, Structured output usually works better than numbered or bulleted lists when you’re calling FDA through CSA. Lists tend to be seen as guidance, so CSA may rewrite them or skip a step, which explains why your Step 3 works every time in FDA but is hit-or-miss in CSA. CSA often treats 'always respond with' as optional formatting and may drop it while shaping the reply. a way to fix this is to make Step 3 part of the required output itself (for ex, step3: this is a AI response) and, if possible, add a simple example. Thinking of the response as a strict format rather than a checklist usually makes things much more consistent.
Thanks.
Hello kevinvu99
In short: your Fabric Data Agent (FDA) behaves consistently when you test it directly because its instructions and tools are applied in a clean, single-layer context. But when you go through Copilot Studio (CS), you introduce a second orchestration layer with its own system prompt, safety & grounding logic, token budgeting, and ranking heuristics. Those layers can dilute or truncate your FDA’s step-by-step instructions, leading to partially executed workflows (e.g., steps 1 and 2 done, 3 skipped).
Below is a practical checklist I use to make CS→FDA integrations far more reliable.
Why CSA→FDA can be less consistent
Prompt layering & instruction collision
- CS has its own system prompt, safety policy, and “helpful assistant” objectives. These can “compete” with your FDA’s instructions or cause CS to shortcut an answer before fully delegating to the FDA.
Context window / token budgeting
- CS often adds summaries, safety annotations, memory, and grounding snippets. Your FDA’s step-by-step guidance may be truncated or pushed out of the top of the context, especially on longer conversations.
Action/orchestration heuristics
- CS tries to be helpful and might answer from its own grounding (Bing/SharePoint/OneLake connectors, etc.) or partial FDA outputs, instead of strictly handing off.
Temperature/variability & tool timeouts
- Default variability and hidden timeouts can cause incomplete runs (e.g., step 3 timing out or being skipped if the FDA returns slower for that step).
Schema-free responses
- If FDA returns natural language only, CS has no reliable way to verify that all steps were executed. It may consider the answer “good enough” and finalize.
Hope this helps, please appreicate giving a Kudos or accepting as a Solution!