Describe the process once. It runs on its own.

Scheduled agents that fire as the person who created them, inside the workspace they were created in, and leave a deliverable behind rather than a transcript.

Described, not drawn

The workflow builder is a conversation.

Automation tools ask you to draw the process before you understand it, then maintain the drawing forever. Say what should happen and when, the way you would tell a colleague, and what runs on Monday is what you agreed on Friday.

  • No graph to maintain The exceptions you did not think to draw are handled on the run, not in a rebuild.
  • Changed in language Correcting the process is a sentence, so the person who owns it can do it.
  • Cron or interval Whatever cadence the work actually has, down to a fixed number of minutes.
One automation: 1. You describe it, What should happen, on what cadence, and what good looks like.; 2. It fires as you, Your access, your workspace. Membership re-checked on every run.; 3. It stops on a ceiling, Iterations, tokens and wall time, set per automation.. The result: A deliverable and a trail, The brief or the updated record, plus every call that produced it..
Owned, not orphaned

An automation is somebody's, and stays that way.

It runs as the person who made it, so it can never reach further than they can. Remove them from a workspace and everything they own stops on the next run, because membership is checked then rather than captured at setup.

  • Runs as a person Not a service account with its own standing permissions to audit separately.
  • Walled to one workspace The same namespace as an ordinary turn. No second path into your systems.
  • Offboarding is not a job Nothing to clean up, and no sync whose failure leaves access behind.
How isolation works →
Three scheduled automations, each with its cadence: a daily standup brief, a weekly churn-risk sweep, and a nightly invoice reconciliation.

What teams schedule

The processes that come up first.

Each of these is a page, and each one started as somebody doing it by hand.

Chase the invoice without being the one who chases

It knows who is fifteen days late and who is the client you do not push. The routine nudge goes out on its own; the delicate one reaches you with the draft already written.

Read the process

Monitoring

Receivables watched by age and by relationship, not by a single threshold.

Aging detection

The fifteen-day case and the sixty-day case are not the same conversation.

Matched follow-up

Routine chases send. The awkward one arrives at your desk, drafted.

Workflow automation

Questions people ask about automations

How is this different from a workflow builder like Zapier or n8n?
Those ask you to draw the process as a graph of steps before you run it, and to maintain that drawing as the process changes. Here you describe the outcome and the rules in language, and the agent works out the steps on each run. The trade runs both ways: a drawn graph is more predictable, and a described process handles the exceptions you did not think to draw.
What does an automation actually produce?
A deliverable rather than a transcript: a written brief, an updated record, a file, a message sent. Alongside it comes the activity trail: which tools were called, with what, and what came back. Both are kept against the workspace and the person the automation runs as.
What happens when the person who created it leaves?
It stops. Membership is re-checked on every run rather than captured when the automation was created, so removing someone from a workspace ends the reach of anything they own on the next scheduled run, with no cleanup job to remember.
How often can an automation run?
On a cron expression or a fixed interval. The practical limit is not the schedule but the ceilings: each automation carries a maximum number of iterations, tokens and wall-clock seconds, so a run that fails to converge stops rather than continuing.
Can an automation reach systems the agent cannot reach in chat?
No, and deliberately. An automation uses the same connectors, under the same credentials, inside the same workspace boundary as an ordinary turn. Your systems have one path into them, which is what lets a reviewer check that path once.
What happens when a run fails halfway?
The trail records what ran before it stopped, so you can see where it got to rather than guessing. Runs that need a person are a design choice rather than a failure. An MCP app is where somebody approves the case and the work continues on the other side.

More of the platform

Each capability has a page. These are the ones nearest this.

See the whole platform

Automate the process that costs you the most.

Start with discovery

Read the docs