Skip to main content

Sentinel

Sentinel is your organization-wide view of cloud agent activity — runs in progress, recent history, and the changes agents produce. It helps teams keep automated work visible and reviewable, not silent background noise.

Think of it like a control room for cloud agents: one place to see what fired, whether it succeeded, and what to open next (logs, diff, or pull request).

Sentinel matters most when you run automations on a schedule or events, or when several people start cloud runs. It complements the per-run view in the Agents workspace with an org-level picture — similar to how Cursor shows run history on the agents dashboard, Devin surfaces activity on automations, and Factory tracks automation impact in Software Factory.

When to use Sentinel

ScenarioWhy open Sentinel
Automations are enabledConfirm triggers behave as intended — not too often, not silently failing
Multiple cloud runs at onceSee what's active without opening each run separately
Something looks wrongFind failed runs, long jobs, or unexpected repos quickly
Budget or usage spikeCorrelate run volume with Usage
Handoff or reviewSee who started a run and what an automation produced

If you only run occasional one-off cloud tasks, the Agents workspace may be enough. Sentinel earns its keep as automation and team usage grow.

In practice

You want to…Sentinel helps by…
Answer "what ran overnight?"Listing recent runs in a time window you choose
Catch a broken automationSurfacing failures and patterns (same repo, same trigger)
Review before mergeLinking runs to diffs, commits, and pull requests
Stop duplicate workShowing repeated runs on the same failure or PR
Report to leadershipSummarizing activity and spend signals without digging run-by-run

You still review changes before they ship. Sentinel makes that review easier to schedule, not optional.

What you can see

ViewWhat it tells you
Active runsWhat's in progress — repository, how long it's been running, what started it
Recent historyCompleted, failed, and stopped runs in the window you select
Changes producedDiffs, commits, and linked pull requests per run
Spend signalsHow automated volume relates to your allowance
Who and what triggeredManual start vs. named automation — accountability without opening every run

Filter by last day, last week, or last month depending on whether you're responding to an incident, doing a weekly automation check, or planning capacity.

Run states

StateIn plain termsWhat to do
WaitingQueued for a workerUsually normal; check back if it stays waiting a long time
In progressAgent is workingWatch scope; cancel from the Agents workspace if it looks wrong
CompletedFinished successfullyReview diff or PR
FailedSetup or task errorOpen logs — often environment, secrets, or network (setup)
StoppedCancelled by someone or policyConfirm that was intentional

Review activity

  1. Scan what's active

    Look for runs that run longer than you expect or target repos you don't recognize. Compare against your automation list.

  2. Open failures first

    Failed runs usually fall into two buckets: environment (install, secrets, network) or task (the agent couldn't complete the instruction). Logs tell you which.

  3. Sanity-check automation output

    For each enabled automation, confirm recent runs match intent — useful output, no duplicate PRs, no changes outside the scoped repository.

  4. Connect volume to usage

    If run count jumps, check Usage. Spikes often trace to a newly enabled automation or a trigger firing more often than expected.

Share a run link

Like Cursor's shared agent URLs, you can send teammates to a specific run so they see the conversation, changes, and outcome — subject to your org's repository access rules.

Watch for

SignalOften means
Same automation failing repeatedlyPrompt, trigger filter, or cloud setup needs a fix
Many runs, little useful outputTrigger too broad or instruction too vague — tighten scope
Sudden run volume spikeNew automation enabled, webhook loop, or schedule firing more than expected
Runs on unexpected reposWrong automation scope or integration permissions — review with an admin
Pair with budgets

Sentinel is most useful alongside Usage & budgets. Visibility plus limits keeps automated spend predictable.

For admins

A short routine keeps cloud agents trustworthy at scale:

  1. Daily (a few minutes) — Scan the last 24 hours; address failures; confirm nothing is stuck in progress unusually long.
  2. Weekly — Review enabled automations: last run time, useful output, disable stale ones.
  3. After a spike — Pause the noisy automation, cancel in-flight runs if needed, fix trigger or prompt, test once manually, then re-enable.
Runaway automation

If an automation fires too often, disable it first to stop new runs, cancel active ones from the Agents workspace, fix the trigger or instruction, then turn it back on after a successful manual test.