Runs
A run is one time an agent was asked to do something. Operate › Runs lists every run in a tenant: which agent ran, how it ended, how many tools and model calls it made, how many tokens it used and how long it took. Open a run to see what it did, step by step.
Use Runs to answer “is this agent behaving, and is it getting more expensive”. Use a trace to answer “what exactly was said at 14:03”.
Who can open it
Section titled “Who can open it”Admin, Builder and Auditor, and platform admins. A Reviewer and an End user do not see Runs.
What you see
Section titled “What you see”The runs list
Section titled “The runs list”- The page header, with a tenant picker when you can see more than one tenant.
- The Recent runs table, newest first, with a search field (Search runs) and two filters:
- Agent: Every agent, or one agent.
- Outcome: Every outcome, Completed, Reached a ceiling, Failed, Waiting on approval or Rejected.
- When the list shows one conversation’s runs, a tag Conversation
<id>above the table. Close the tag to show every conversation.
| Column | What it shows |
|---|---|
| Run | The agent’s name and when the run started, as a link to the run. Under it: when, relative; Through <assistant> if it came through a published assistant; Continues after an approval if an approval resumed it; and the first eight characters of the run id. |
| Outcome | How the run ended (see below). |
| Tool calls | Tools the agent called. |
| Model calls | Calls to the model. |
| Tokens | Tokens the run used. |
| Took | How long the run took. |
| Asked by | The person who asked, when the run records one. |
| Trace | A link that opens the run’s trace beside the list. Shown only when tracing is on and the run has a trace. |
Searching asks the server, so it reaches every run in the tenant, not only the ones on screen. The tenant, the filters and the search are kept in the page address, so a run you open comes back to the same list, and a link can name one agent’s runs.
When no run has been recorded yet, the page says No runs yet and offers Choose an agent. When the filters match nothing, it says No runs match these filters and offers Clear filters.
Outcomes
Section titled “Outcomes”| Outcome | Meaning |
|---|---|
| Completed | The agent answered. |
| Reached a ceiling | The run stopped at one of its limits: tool calls, model calls or tokens. The tag names which. The answer so far is kept. |
| Failed | The run ended in an error. Open its trace for the reason. |
| Waiting on approval | The agent asked for a tool that needs a person’s approval, and is waiting. See Approvals. |
| Rejected | A person refused the held tool call, so it did not run. The tag names the tool. |
A run’s page
Section titled “A run’s page”Select a run to open it. The header shows the agent’s name and, when the run has a trace, View trace, which opens the trace beside the page.
The run lists the facts: Run id, Agent (a link to the agent), Outcome, When (exact and relative), Took, Asked by, Tool calls, Model calls, Tokens, and, when they apply:
- Cost, when the run’s spend was priced.
- Conversation, a link to every run in the same conversation.
- Through the assistant, a link to the assistant the run came through.
- Continues, a link to the run an approval resumed.
Approvals, shown when the run held tool calls for approval: each held tool, its state (Waiting, Approved, Rejected or Expired), when a waiting one expires, who answered and why, a link to the Resumed run, and Review the approval for <tool>.
What it did: the run’s steps in order, as a tree. Each step says what the agent did, how long it took and what it used:
| Step | Meaning |
|---|---|
| asked the model | A model call, with the model and its tokens in and out. |
| called | A tool call, with the size of its result. |
| read | A knowledge lookup, with the number of passages and documents. |
| delegated to | A hand-off to another agent, whose own steps nest under it. |
| needs approval for | A tool call held for a person. It did not run. A link opens its approval. |
| was refused | A tool call the platform refused, with the reason. |
| answered | The final answer. |
Runs recorded before step recording existed show only their totals, and say so.
What you can do
Section titled “What you can do”See one agent’s runs
Section titled “See one agent’s runs”- In Agent, choose the agent.
- Or open the agent’s page in Build › Agents and follow its link to its runs.
See every run in a conversation
Section titled “See every run in a conversation”- Open a run that belongs to a conversation.
- Select the link beside Conversation.
The list shows that conversation’s runs, with a tag naming it.
Find out why a run failed
Section titled “Find out why a run failed”- Filter Outcome to Failed.
- Select Trace on the run, or open the run and select View trace.
If the run has no trace, tracing is off for the platform; ask your platform administrator.
Answer an approval a run is waiting on
Section titled “Answer an approval a run is waiting on”Open the run, then select Review the approval for <tool> under Approvals. Answering takes the Admin or Reviewer role; see Approvals.
Limits
Section titled “Limits”| Limit | Value |
|---|---|
| Runs listed | The 200 most recent per request; search and filters reach older ones |
| Cost | Runs record tokens. A money figure appears only when the run’s spend was priced |
How long runs are kept is set per tenant; see Data retention.