Runs and approvals

Follow agent runs, approve or reject the tool calls they wait for, stop a run, read its trace, and fix a failed run.

Every run of every agent in the workspace is on the Runs page, and every tool call waiting for a person is on the Approvals page. Find a run# In Agents, select Runs. Select a row to open the run. The list shows the Agent, Status, Kind (Test, Invoke or Evaluation), the published Version (none for a test) and when it Started. It refreshes every 15 seconds. Read a run# The top of the run page shows: Item Meaning Status Pending, Running, Approval required, Cancellation requested, Completed, Failed or Canceled. Step N of M Model calls made, against the agent's step limit. Tokens Prompt and completion tokens of all model calls. Cost Computed from the model's price when the version was frozen, or No price. Below it are the Output (when the run has finished), the Steps and the Inputs. Each step has an icon for its kind: a model call, a vector store search, a tool call or an approval. Select a step to see its input, output and details. The page updates live. If the live connection drops, it says Connection lost. Refreshing every few seconds. Approve or reject a tool call# When the agent wants to call a tool with Approve each call on, the run stops with the status Approval required and a Review approval button. You need tools.edit in the workspace. Select Review approval, or select Approvals in the Agents menu. Read the call: the tool name, what it points to (for example the package Invoice.Processing) and the exact arguments the agent wants to pass. Run of links to the run; Expires shows when the request lapses. Select Approve call, or Reject call and confirm with Reject call. After an approval the agent makes the call with exactly those arguments. After a rejection the agent is told the call was not approved and carries on without it. A request that nobody decides before it expires, or that is still open when the run ends, expires. The run shows the decision as a step, followed by the tool call. The Approvals page lists the pending calls of all agents in the workspace and refreshes every 10 seconds. When there are none it says No approvals waiting. Stop a run# You need jobs.cancel. On the run page, select Stop run. Confirm with Stop run. The status is Cancellation requested until the Robot confirms. The agent stops after its current step; a tool call already sent may still complete. Read the execution trace# A run is also a job of the Agent Runner process, with the same ID. Its full trace is available from the Orchestrator: In the workspace, open Automations ▸ Jobs. See Jobs. On the run's job, select ⋮ ▸ View traces. The Execution trace page lists every span with its Status, Duration, Start, End and Tags. The agent.run span covers the whole run; each step is a span under it. Traces are kept for 30 days by default (configuration key Agentic:TraceRetentionDays, up to 365). When a run fails# A failed run shows one line that says what went wrong. Some lines come with Open models. Message What to do Budget reached for this month. The AI budget is used up. Ask an administrator to raise it, or wait for the next month. The workspace's AI policy refused this call. The model is not allowed by the workspace's AI policy. Choose another model. The model provider did not answer. Run the agent again. A temporary provider problem. Run again. The model refused the request. Check the agent's model. Check the model binding and the connection on the Models page. The answer did not match the output schema. Adjust the instructions or the schema. Make the instructions ask for exactly the fields of the output schema, or relax the schema. A tool call may have run. Check its target before running again. The outcome of a tool call is unknown. Check the system it changes before you run again, so the change is not made twice. The run called something its version does not list. Test the agent again. Test and publish a new version. A tool call ran without an approval. A tool that needs approval was called without one. Test the agent again. A run that stays Pending is waiting for a Robot with Python in the workspace. See Machines and robots.

Find a run

Read a run

Approve or reject a tool call

Stop a run

Read the execution trace

When a run fails