Sessions & Runs
One page for everything that ran. Watch live, replay history, and unblock runs waiting on a human.
Every execution in your workspace (a manual launch, a scheduled run, an email-triggered job, an API call) lands on the Sessions page. It's the operational home of the platform: what's running now, what finished, what failed, and what's waiting on a person.
Workflows and Agents
The page has two tabs:
- Workflows shows sessions from your browser workflows: every authoring session and every execution of a published workflow, newest first.
- Agents shows runs of your agent automations from the Agents page. Each run carries a Sandbox or Live tag, so test runs never blend into production history. You can rehearse an automation all afternoon in sandbox and still see at a glance which runs actually touched real systems.
Both tabs share the same filters, the same statuses, and the same actions, so operating an agent automation feels identical to operating a workflow.
Finding a run
Above the table, filter pills split the list into All runs, Needs input, Completed, and Errors, with live counts on each pill. The counts alone are a health check: a growing Needs-input count means people are the bottleneck; a nonzero Errors count is your morning to-do list.
Combine the pills with the search box and the per-workflow filter to answer questions like "show me every failed run of the invoice loader this week" in two clicks. Refresh pulls the latest without reloading the page.
Statuses you'll see
Watching a live run
Click a live session and the workbench opens in live view: the actual browser, working in real time. You see every page it visits, every click it makes, and every step it narrates in the chat alongside. This isn't a status bar pretending to be progress; it's the same view you had when you taught the workflow in the first place.
Live view is also interactive. If the run pauses to ask something (an ambiguous choice, an unexpected page, a value it can't find), you can answer right there in the chat and let it continue. For a run you're not sure about, watching the first execution live is the cheapest insurance there is.
Terminating a run
Need to stop a run? Terminate ends a live session immediately. The session is kept with everything it did up to that point, marked Terminated, so nothing disappears: you can still replay what happened before you pulled the cord. Terminate is for "stop touching that system right now"; it doesn't delete anything.
Replaying a finished run
Click Replay on any finished session and the workbench opens on its History tab: the session recording plus a read-only session log covering what was said, which steps ran, and the final output. It's the fastest way to answer "what exactly happened on Tuesday's run?" without guesswork.
The difference between Watch and Replay is simply tense. Watch is the present: a live browser you can talk to. Replay is the past: a faithful record you can scrub through, but not change. Both show the same run from the same seat.
The History tab has more to offer, including step-by-step traces and one-click fixes when a run fails. That's covered in Phase 3: Test & Iterate.
The Needs-input queue
Workflows can include human-in-the-loop checkpoints: an approval before submitting, a judgment call the agent shouldn't make alone, a value only a person knows. When a run reaches one, it shows up under Needs input and waits.
Treat that pill as a queue. Anyone on the team can open a waiting run, see exactly what it's asking and the context it's asking in, complete the step, and the run picks up where it left off. Nothing times out silently and nothing proceeds without the answer; the run is simply parked until a person weighs in. Teams that run approval-heavy workflows check the Needs-input count the way they check a shared inbox.
Support, scoped to the session
Every session row has a Support action that opens a support ticket scoped to that exact session. The team helping you sees the same run you're looking at (the recording, the log, the inputs), so there's no copy-pasting IDs into emails and no "can you describe what happened?" round trips. If a run did something surprising, Support on that row is the shortest path to an answer.
Where to next
The four ways a run starts: manual, API, email, and schedule.
Simulation, the History tab in depth, and one-click fixes for failed runs.
Was this page helpful?