Agents overview
An agent is a standing instruction. You write a prompt once, say when it should run and what it may touch, and Corte runs it without you — on a schedule, or the moment your content changes.
Where a workflow is a pipeline you press Run on, an agent is something that notices. "Every Monday at nine, summarise everything we published last week and draft next week's teasers." "Whenever a product page is published, check its images have alt text and fix the ones that don't."
When to use an agent
| Reach for… | When… |
|---|---|
| Studio | You want one image or video, right now. |
| Workflows | The job has steps you'll repeat, and you'll press Run. |
| Agents | The job should happen without you — on a clock, or in response to a change. |
An agent can use Corte's own tools, so the two overlap: an agent with the right access can do most of what you'd do by hand, including running a workflow. The question isn't which is more powerful, it's whether something needs to happen when nobody is looking.
How a run works
Each run executes in a fresh, isolated sandbox: a private container that starts empty, and is destroyed when the run finishes. Nothing carries from one run to the next — no files, no memory of last time — so a run that goes wrong leaves nothing behind for the next one to trip over. Inside, the agent has ordinary file and shell tools for its own scratch work, plus whichever Corte tools your access choices unlock. The only credentials in there are scoped to that one run and stop working the moment it ends, so nothing it holds can be used later or elsewhere.
It acts as you, which is what makes the access step matter: anything it creates belongs to you and lands where your own work does, in the Library and in Content.
The Agents page
Open Agents from the rail:
- New agent — the three-step wizard. See Creating an agent.
- Your agents — each one shows when it runs and what it can reach. The toggle turns an agent off without deleting it, which is the right move when something is misbehaving: a disabled agent has no schedule and ignores content changes until you turn it back on.
- Opening an agent shows its runs — what happened, when, and what it cost. See Runs and costs.
What agents cost
Creating, editing and disabling agents is free. A run costs what the work costs: the model tokens the agent spends thinking, plus the time its sandbox is awake, plus any generation it performs.
All of it settles as one line per run in your usage history, so a run reads as a single number rather than as scattered charges you have to add up. See How credits work and Usage and ledger.
Two limits keep an agent that goes wrong from being expensive: how many runs you can have going at once, and how many can start in an hour. Both depend on your plan, and a run stopped by either says so in the feed rather than vanishing — see Runs and costs.
Keep going
- Creating an agent — the wizard, triggers, and choosing what it may touch.
- Runs and costs — reading the feed, cancelling, and what to check when nothing ran.