Hello Mahadevanp 

Unfortunately there is currently no way to reserve or allocate a capacity to be used for Data Agent only. Fabric capacity units are shared across all items of a workspace. 

What would be appropriate for you is to move all items pertaining to the data agent to another workapoace assigned to a different capacity. This will ensure that when total capacity exceeds throttling will be experienced on a smaller footprint, and it won't impact other workloads.

One point to remember, data agent is considered a background workload, and as such is smoothed over a 24 hr period. Due to smoothing, only a portion of the CU usage for an operation applies to any individual timepoint, which reduces throttling overall.

More details here:

Understand your Fabric capacity throttling - Microsoft Fabric | Microsoft Learn

Hi Mahadevanp 

No — not currently. As deborshi_nag mentioned, there is no native mechanism to carve out a dedicated, isolated capacity pool exclusively for the Data Agent. It shares the workspace's Fabric capacity (CUs) along with everything else.

However, there are a few other practical strategies (along with one mentioned above) to mitigate the risk of Data Agent queries crowding out other workloads :

1. Use Autoscale on Your Capacity
Enable autoscale so the capacity can temporarily exceed its base CU limit during bursts rather than hard-stopping. This won't isolate the agent, but it prevents the "all operations will shut down" scenario during spike periods. Note: autoscale adds cost.
2. Enable Surge Protection (Admin Setting)
In the Fabric Admin portal, you can configure Surge Protection on the capacity. This prevents any single workload from consuming all available CUs by rejecting background jobs when a recovery threshold is breached — giving your interactive and higher-priority jobs a fighting chance. Data Agent, being a background job, would be among the first rejected under surge pressure, which is actually somewhat protective of your other workloads.
3. Understand the Background Job Smoothing Benefit
The Data Agent is classified as a background operation, which actually gives it a more forgiving consumption profile than it appears. The smoothing policy means each ~6.67 CU-minute request only "consumes" 1 CU-minute of capacity per hour. On an F64, this equates to roughly 13,800+ agent queries per day before exhaustion — so for many use cases, the shared pool isn't as constrained as it seems.

Another important Aspect to check

How to Monitor Data Agent Consumption — Clear Step-by-Step

The Fabric Capacity Metrics App is your go-to tool. Here's exactly where to find Data Agent usage:

Step 1 — Install/Open the Metrics App Go to AppSource or your Fabric Admin portal and open the Microsoft Fabric Capacity Metrics app.

Step 2 — Select Your Capacity At the top of the Compute page, use the Capacity Name dropdown to select the capacity where your Data Agent lives.

Step 3 — Find Data Agent in the Matrix Table Scroll to the Matrix by Item and Operation table at the bottom. Filter by:

• Item Kind = LlmPlugin ← This is the Data Agent's item type
• Operation = AI Query ← This is the operation name for both input prompts and output completions

Step 4 — Columns to pay attention to:

• CU(s) — sort descending to see who's eating the most capacity
• Users — how many people are actively querying the agent
• Rejected count (add this optional column) — if you're seeing rejections, that's a red flag
• Overloaded minutes (add this too) — tells you how long the capacity was under strain

Step 5 — Check Throttling Chart Switch to the Throttling tab in the Capacity Utilization visual. Since Data Agent is a background job, watch the Background Rejection threshold (triggered when 24-hour background smoothing exceeds capacity) — this is when Data Agent queries will start being dropped.

Step 6 — Drill Into a Specific Timepoint Click any column in the Utilization or Overages chart → hit Explore to drill into the Timepoint Detail page and see exactly which LlmPlugin items fired at that moment.

Quick Cost Reference:

• Input tokens: 100 CU seconds per 1,000 tokens
• Output tokens: 400 CU seconds per 1,000 tokens
• A typical query (2,000 input + 500 output tokens) costs ~6.67 CU minutes

Some helpful link for more details:

• Data agent consumption - Microsoft Fabric
• Fabric operations - Microsoft Fabric
• Understand the metrics app compute page - Microsoft Fabric