Perform Web Task
You pass a single natural-language instruction to an agent on the browser. The agent performs that one instruction — navigating, clicking, extracting — and returns the result. It does not manage a larger flow; you decide when to call it and what to do with the result. Use it for ad hoc, exploratory, or frequently changing flows where you don’t need the same result every time. Call it withagent.task in the SDK or perform-web-task over REST. Supports async execution.
Automation Tasks
You define a task by writing a prompt or demonstrating it, and Anchor compiles it into deterministic code and a reusable workflow. Runs repeat reliably and produce consistent output. If the deterministic code breaks — for example when a page changes — the workflow falls back to AI to complete that step, and Anchor can self-heal the workflow afterward. Use it for repeated workflows that need to run the same way every time — production, CI, or anything you deploy and reuse. Tasks can be versioned, edited by hand, and shared with end users as links. Code Tasks — write TypeScript yourself or hand-edit workflow JSON when you need direct control over the automation code.Comparison
| Perform Web Task | Automation Tasks | |
|---|---|---|
| You provide | A natural-language instruction, per call | A prompt or a demonstration |
| Flow control | You decide each call; agent executes one instruction | Defined by the workflow |
| Behavior | Varies run to run | Deterministic, consistent across runs |
| Best for | Ad hoc, exploratory, changing flows | Repeated, reliable, cost heavy production runs |