Preview and test an agent

Preview is a conversation with one agent, with its steps shown beside the answer as they happen. The Run section on an agent’s page sends a single message. Both run the saved agent for real: it is not a sandbox.

Every turn is a real run

A preview turn and a Run are metered against the tenant’s budget, written to the audit trail, and call the tools the agent is granted. A write tool that is not held for approval does its work. Hold write tools while you test; see Create and edit an agent.

Who can open it

Role	What they can do
Admin	Open Preview, send messages, use Run, publish the agent.
Builder	The same as an Admin.
Auditor	Cannot open Preview and does not see Run. Reads the resulting runs in Runs.
Reviewer, End user	Cannot open it. A Reviewer answers the approvals a preview raises.
Platform admin	Everything an Admin can, in any tenant.

What you see

The Preview page

Open it with Preview on the agent’s page or on its row in Build › Agents.

1. The header: Preview, with New conversation (available once a turn exists) and Publish as assistant.
2. Conversation, on one side:
   • Before the first message: Ask the agent something. Its steps appear beside the answer as they happen.
   • Each turn as You: and Agent:, and under each answer a line with Model calls, Tool calls, Tokens and Took, and View this run.
   • Waiting on a person, when the agent asked for a tool that needs approval, with a Review button per held tool.
   • Message, and Send. Ctrl+Enter (⌘+Enter on a Mac) also sends.
3. What it is doing, on the other side: the steps of the current turn, in order (see Reading the steps). Below them, the conversation’s short id and turn count, with Its runs, a link to every run in this conversation.

If the agent’s details cannot be read, the page says so and you can still talk to it: a run uses what is saved on the server.

The Run section

On an agent’s page, Run holds a Message box and Run. It sends one message and shows:

• the answer, or The run failed with the reason;
• Answered without a gated tool when the agent asked for a tool that needs approval: the answer was produced without it, and Review the approval for the tool opens the approval;
• the steps, and the line Tool calls made inside a collaborator, not listed above when a collaborator called tools;
• Model calls, Tool calls, Tokens and Took;
• View this run and View its trace.

What you can do

Hold a conversation with an agent

1. Open the agent and choose Preview.
2. Type a message in Message and choose Send.
3. Watch What it is doing fill in while Working… shows.
4. Send the next message. Every turn continues the same conversation.
5. To start over, choose New conversation.

Test a tool that needs approval

1. Send a message that makes the agent call a held tool.
2. The conversation shows Waiting on a person: The agent asked for the tool**, which needs approval.**
3. Choose Review to open the approval, or ask an Admin or a Reviewer to answer it in Approvals.
4. Approving runs the tool as a new run, linked to the first; the conversation continues here. Rejecting runs nothing and records the reviewer’s reason.

An approval left unanswered for 24 hours expires and the tool never runs.

Reading the steps

Each step names who acted and what they did, with a status: ok, running, awaiting approval, failed or not started.

Step	Meaning
asked the model	A model call, with the model and tokens in and out.
read	A retrieval from the agent’s collections: Passages and documents.
called	A tool call, with the size of its result; (truncated to the bound) when the result was cut to the platform’s limit.
delegated to	A hand-off to a collaborator. Its steps follow, indented.
needs approval for	A held tool call.
was refused	A call the platform refused, with the Reason.
answered	The final answer.

Some steps unfold to show their Input and Output. When the agent delegates, a line at the end reads Tokens the supervisor spent deciding where to route, so routing cost is visible apart from the answer.

Publish from the preview

Choose Publish as assistant. The assistant form starts from this agent; see Publish and share.

Limits

Limit	Value
A run	Ends after 5 minutes, even if you leave the page.
One tool call	60 seconds by default, set by the platform administrator. A call that runs out of time returns an error to the model and the run goes on.
One tool result	32 KB by default, set by the platform administrator. A longer result is cut and the model is told.
Per-run ceilings	The agent’s Tool calls per run, Model calls per run and Tokens per run (defaults 20, 25 and 200,000). A run that reaches one stops and says which; the answer so far is kept.
Approval window	24 hours.
Message	Required; an empty message is refused.

When it does not work

What you see	What to do
Running an agent takes the Admin or Builder role	Ask an Admin for the Builder role.
The run failed	Read the reason. A model that is no longer deployed, a server out of service, or a refused delegation each fails the run and says so.
A run outcome of Reached its tool call ceiling (or model call, or token)	Raise that ceiling on the Edit agent page, or narrow the task.
Answered without a gated tool	Expected while the call waits; answer the approval.

Related

• Agents
• Create and edit an agent
• Collaborators
• Approvals and Human approval
• Runs and Traces
• Playground, for trying a model without an agent