Running Agents

Once an agent has a model — and any tools or memory it needs — you can run it with a task and watch it work in real time.

Starting a Run

In the agent editor, click Run in the header. If you have unsaved changes, they’re saved first; the right panel then switches to Run Agent.

Type a task or question — for example “Get the weather in Istanbul and summarize in 3 bullets.” — then press Ctrl+Enter (Cmd+Enter on Mac) or click Run Agent.

A full-page version of the same screen is also available by appending /run to the agent’s editor URL (/agents/<agent-id>/run). It shows the agent’s name, model, and status in a header, with the same prompt box (“What can I help you with?”) above the run monitor.

The Run Monitor

The run monitor shows a running transcript:

Your message, under You
The agent’s response, under Agent, with a badge showing the model it used (Provider / model, for example OpenAI / gpt-4o)

While the agent is working, a thinking indicator shows the current phase: Initializing…, Planning next action…, Executing action…, Analyzing results…, Deciding next step…, or Working….

Tool calls

Each tool the agent calls appears as a card with its action name, the piece it belongs to, and — when recognized — a short summary of the key argument (such as a location, channel, search query, or message). A status icon shows whether the call is still running, succeeded, or failed. Click a card to expand its Parameters, Result, and, on failure, Error.

Final answer and output

The agent’s final answer renders as Markdown.
If the agent has an Output Parser, the parsed result appears below it as formatted JSON with a Copy button.
If the run fails, the error message appears in its own block.
Once the run finishes (completed or failed), a footer shows the step count, tool call count, and total duration.

A connection indicator near the input shows whether the live stream is connected.

Run Status

Status	Shown as	Meaning
`pending`	No badge yet	Run created, not yet started
`running`	Running	The agent is actively working
`waiting_approval`	Waiting for approval	The run is paused for an approval decision
`waiting_input`	Waiting for input	The run is paused for more information from you
`completed`	Completed	The agent finished and returned a final answer
`failed`	Failed	The run stopped because of an error or an exceeded limit
`cancelled`	Cancelled	You stopped the run

Approval and Input Requests

If a run is Waiting for approval, an Approval Required dialog opens over the monitor, showing a risk badge (Low, Medium, or High), the proposed action’s description, the tool name, and its details as JSON.

Approve — optionally add a note in Response (Optional) first.
Reject — switches the dialog to a required Reason for Rejection field; confirm with Confirm Rejection.

If a run is Waiting for input, an Input Required dialog shows the agent’s question under Agent’s Question, with a Your Response box to type your answer (Ctrl+Enter submits) and a Submit button.

The run resumes automatically once you respond.

Cancelling a Run

While a run is active, click Stop in the header to cancel it — you’ll see a confirmation: “Agent run has been cancelled.” Stop is only shown while the run is active; it’s replaced once the run finishes.

Starting Over

Once a run ends, New Chat (full-page view) or Reset (editor run panel) clears the transcript so you can send a new task to the same agent.

Limits and Credits

Every run has built-in limits that aren’t configurable from the editor: 50 steps, 100 tool calls, and a 5-minute total time budget. A run that hits any of these stops automatically and fails with a budget-exceeded error.

Each planning step calls the agent’s selected AI model using your puq.ai balance, the same as other AI features on the platform. If your balance runs out mid-run, the run fails. See Billing.

Best Practices

Keep tasks specific — the clearer the instruction, the fewer planning steps the agent needs.
Watch the tool call cards during a run to catch a wrong tool choice early; cancel and adjust the system prompt or tools if it’s heading the wrong way.
If a run keeps hitting the step or tool-call limit, the task is probably too broad for a single run — split it into smaller agents or steps.