Skip to content

Jobs and monitoring

The Jobs section of the console shows every agent run for your workspace: past, in-flight, and failed.

The jobs list shows one row per job, newest first, with a Load more button for older jobs. Filter by status, agent, account or project; the filters live in the page URL, so a filtered view is linkable.

Column Description
Status One badge per job: the status while the job is in flight, the outcome once it has finished (see below)
Agent The agent name that drove the job
Triggered by The GitLab user whose action fired the trigger, and the managed account the job ran as
Project The project path, with a link to the issue or MR
Event The GitLab event that fired the trigger
Started When the job was dispatched
Duration Elapsed time from start to finish
Status Meaning
PENDING Queued, waiting for a free mate
DISPATCHING A slot was found and the job is being placed
STARTING Container is being allocated and the agent is starting up
RUNNING Agent is actively executing
FINISHING Job received a stop signal and is winding down
DONE Job completed normally
CANCELED Job was stopped by a user action, a timeout, or a policy limit
ERRORED Job ended unexpectedly. Check the log for details

A job has an outcome once it reaches a terminal status. The outcome replaces the status in the badge; non-SUCCESS outcomes render red:

  • SUCCESS — the agent exited cleanly.
  • FAILURE — the agent exited with an error, or the job ended without a clean exit.
  • CANCELED — the job was stopped by a user action or a policy limit (admin stop, unassignment, etc.).
  • TIMEOUT — the job was stopped because it exceeded the configured wallclock limit.

While any job without an outcome is visible, the list refreshes every 5 seconds. If the refresh fails (for example, due to a network interruption), a warning banner appears. Rows that already have an outcome are not re-fetched automatically.

Click any row to open the job detail view.

The job detail view shows the run log for a single job, including live streaming while the job is running.

The top of the detail view shows:

  • Status badge (the outcome once the job has finished)
  • Triggered by, Managed account, Project, Issue / MR, Agent and Trigger
  • Started, Finished and Duration
  • Tool calls — total count of MCP tool calls
  • Cost — an estimate of LLM spend for the job in USD (e.g. ≈ $0.0124), computed at list price by the agent CLI. Your provider’s invoice is authoritative.

The log lists the run in chronological order (oldest first). Entries are grouped by kind:

Entry What it shows
Agent text What the agent wrote
Thinking The agent’s reasoning, collapsed by default
Tool call Tool name, a one-line summary and duration; click to expand arguments, result or error. Failed calls are red
Container build The devcontainer build, with its base image, state and output
Status change A divider for each lifecycle transition, with an optional detail message
Final status Job outcome, exit code, and an optional summary message

If a run has more events than the console loads, a banner says only the first events are shown.

While a job is in flight, the log is live: new events stream in as they happen and a LIVE badge appears next to the heading. When the final status arrives, the stream closes automatically and the summary card refreshes.

You can pause and resume the live stream using the Pause / Resume button in the summary card. While paused, new events are held back and appear when you resume.

The Cancel job button appears in the summary card while the job has no outcome and is not yet FINISHING. It requires the manager role or higher. Clicking it sends a stop signal to the running agent. The job transitions to FINISHING while it winds down, then ends as CANCELED.