Artifacts
Files your agent produces during a run.
An artifact is a file the agent publishes while it works: a generated report, a processed CSV, a rendered diagram. Artifacts attach to the run that produced them, persist after the run finishes, and download cleanly through either the dashboard or the API.
See them on a run
Open any run from the agent's Runs tab, then switch to the Files tab on the run detail page. Every file the run published appears there with its name, size, and a direct download link. Runs that produced no files show an explicit empty state — the tab is always present.
For sessions that span multiple runs, the session view rolls every file up in one place so you don't have to click into each run individually.
How an agent publishes one
Publishing is a one-line call from inside the agent's recipe — point at a file path and the platform captures it. There's no setup beyond turning the feature on for the node that should be able to publish, and partial writes never reach the run's artifact store: a file only ships once the publish call returns successfully.
The Recipes guide covers wiring up node behaviour; Sandboxes explains where the agent's working files live during a run.
Agents can read them back
Publishing isn't one-way. An agent with artifacts enabled can list what it published in earlier runs and download any of those files into its working directory — so a dataset produced last week doesn't have to be rebuilt to be used today. A fresh run still starts from a clean workspace; recovery is an explicit tool call, never an automatic restore.
Read-back is workspace-scoped: an agent can also list another workspace agent's artifacts by that agent's slug. That's the simplest way to hand a file from one agent to another — the producer publishes it and states the artifact's name in its result, and the consumer fetches it. Note the flip side: any agent in the workspace can read any workspace artifact, so treat artifacts as visible to every agent you run, and keep workspaces as your isolation boundary for sensitive output.
When to use artifacts vs. run output
The run's response payload is the right place for a structured result the caller needs immediately — a classification, a status, a small JSON blob. Artifacts are the right place for anything you don't want to inline: large files, binary content, reports a human will download later, or outputs that accumulate across a long-running session.
From the API
Every dashboard action has an API equivalent: list files on a run, mint a short-lived download URL, or roll up files across a session. The API uses the term "artifacts" for these endpoints — the Runs API reference lists them with schemas and examples.
Download URLs are time-limited by design — request a fresh one whenever you need access. There's no permanent download endpoint, which keeps artifact access auditable and revocable.
Next
- See the Runs guide for how artifacts fit into the wider run lifecycle.
- The Recipes guide covers how to turn publishing on for a node in your agent.