Ask your agent about activity
Read your workspace activity from your agent: what the analytics tool answers, prompts that work, and whose numbers come back.
Your agent reads your activity itself. Ask it a question in plain language and it calls the analytics tool, answering from the same records the Activity page shows.
Before you start
- Your agent is connected to the workspace. See Connect your agent.
- At least one role has run once. An empty workspace has nothing to report.
Nothing to install and no skill to learn: analytics is listed to every member of the workspace, admin or not.
What your agent can read
| Ask about | What comes back |
|---|---|
| The whole picture | Call outcomes, session counts, value returned, guardrail decisions, top skills |
| Value | The total, the uses behind it, the top skills and users, a daily trend |
| One figure, grouped | Calls, failures, value, unpriced uses or sessions, grouped by skill, tool, user, role, status, client or day |
| Sessions | The episodes your agents ran, then every call inside one of them |
| Guardrail decisions | How many calls a rule judged, refused, would have refused in observing mode, modified or errored on, grouped by outcome or by rule |
| Calls | Individual tool calls with the guardrail decision on each, then one call in full: its input, its output and the whole guardrail verdict |
Every question covers a period: the last 24 hours, 7, 30 or 90 days, everything, or two exact dates. Unstated, it is 30 days for figures and 7 days for lists.
Answers start small, at the top 10 or 20 rows. Ask for more and your agent raises the limit or pages through the rest.
Start with the whole picture
Name the period and the figures you want.
How did our workspace do over the last 7 days? Calls, failures and value.Group the figure that matters
One figure, one grouping, one period.
Break our calls down by day over the last 30 days.
Which tools failed the most this month?
Which skills returned the most value over the last 90 days?
Is anyone still using the Refund Policy skill? Check the last 90 days.Open the record behind the number
Go from a figure to the sessions and calls that produced it. Only the single-call read carries payloads, so ask for one call when you want its input and output.
List my sessions with errors from yesterday, then walk me through the worst one.
Open the last failed execute_script call and show me its input and output.
Why was my last call refused?Check it worked
Ask for last week's totals, then open Activity and select the same period. The figures match. Your question is recorded too: it appears in Call Log as an analytics call.
Whose numbers come back
Your permissions decide, never the prompt. Every answer states whose activity it covers and names anything it left out.
- An admin gets the whole workspace.
- A member gets their own sessions and calls. Figures about other people are withheld, with the reason said out loud, and a colleague's session or call answers "not found".
- Grouping a figure by user needs workspace-wide access, so a member is refused that one breakdown and gets their own totals instead.
No prompt can name another workspace. The answer always covers the workspace the agent is connected to, and a member of two workspaces connects twice.
Two things stay out of reach. analytics never writes, and it returns records of what happened, never definitions: your roles, skills and rules are not part of it. A guardrail verdict on a call is a record, so your agent can tell you why a call was refused, but not what a rule checks.
Next
- Read the dependency map, to see how roles, skills and tools connect.
- Act on insights, to turn what you found into a fix.