Skip to main content
The MAS Stats page gives you a detailed breakdown of how each multi-agent system is performing: which agents run most often, which tools take the longest, which models are being called, and how handoffs flow between agents. While the Dashboard shows high-level metrics across your whole operation, MAS Stats zooms into a single MAS to help you optimize its performance.

​
Accessing MAS Stats

  1. Go to Multi-Agent Systems in the sidebar.
  2. Click on a MAS to open the editor.
  3. Switch to the Stats tab.
  4. Select a date range to view metrics for that period.

​
Summary cards

At the top, four summary cards give you the big picture:
MetricDescription
Total ExecutionsHow many times this MAS ran in the selected period.
Total DurationCumulative time spent across all executions.
Avg. DurationAverage time per execution. This is what your customers experience as response time.
Total LLM CallsTotal number of LLM calls across all executions. More calls generally means longer response times.

​
Detailed charts

​
Function (tool) metrics

Shows each tool used by agents in this MAS:
ColumnDescription
NameThe tool name (e.g., Get Customer Orders, Refund Order, your API tools, MCP tools).
CountHow many times this tool was called.
Avg. TimeAverage execution time per call.
This is the most important chart for optimizing response times. If a specific tool is slow (e.g., an external API tool averaging 3+ seconds), it’s dragging down your overall response time. Consider:
  • Checking if the external API can be optimized.
  • Reducing how often the tool is called (via agent instructions).
  • Removing the tool from agents that don’t strictly need it.

​
Agent metrics

Shows each agent in the MAS:
ColumnDescription
NameThe agent name.
CountHow many times this agent was activated.
Avg. TimeAverage time this agent spent active per execution.
High counts on a Triage Agent with low times are normal: it’s doing lightweight routing. High times on a specialist agent might indicate it’s using a model that’s too powerful for the task, or it’s calling too many tools.

​
Response (LLM) metrics

Shows metrics grouped by model:
ColumnDescription
ModelThe model used (e.g., gpt-4.1-mini, gpt-4.1, o4-mini).
CountNumber of LLM calls using this model.
Avg. TimeAverage time per LLM call.
Reasoning models will show significantly higher average times than non-reasoning models. If you see a reasoning model with high count and high average time, consider whether those agents truly need reasoning capability, or if a faster model would produce similar results.

​
Handoff metrics

Shows how agents delegate to each other:
ColumnDescription
From AgentThe agent that initiated the handoff.
To AgentThe agent that received the handoff.
CountHow many times this handoff occurred.
This tells you how conversations flow through your MAS. Look for:
  • Unexpected handoffs: If an agent is handing off to an agent it shouldn’t, the handoff descriptions may need refinement.
  • Missing handoffs: If an agent never hands off to a specialist you created, the handoff description may not be descriptive enough.
  • Dominant paths: If 90% of handoffs go to one agent, you might optimize that agent more aggressively since it handles the bulk of the work.

​
Using stats to optimize

Reduce response time:
  1. Check Function metrics for slow tools. Fix or remove them.
  2. Check Response metrics for slow models. Downgrade where reasoning isn’t needed.
  3. Check Agent metrics for agents that take too long. Simplify their instructions or reduce their tool count.
  4. Enable parallel tool calls in Model Settings for agents that fetch multiple independent pieces of data.
Reduce cost:
  1. Check Response metrics for expensive model usage. Switch agents to cheaper models where quality holds up.
  2. Check Function metrics for high-count tools. Are agents calling tools unnecessarily? Tighten instructions.
Improve quality:
  1. Check Handoff metrics for unexpected routing. Refine handoff descriptions.
  2. Cross-reference with Traces for specific executions that performed poorly.
  3. Create Reports for issues you find.