Giving a Step an AI Agent Instead of a Fixed Action
A workflow step can now be a bounded AI agent — with a goal, an allowed tool list, a step limit, and a dollar budget — instead of a single fixed action.
Most "AI step" features in automation tools are a single prompt box: give it text, get text back. Workflows now supports something more bounded and more capable — a step that's a full agent, with real limits on what it's allowed to do.
What an agent step is
Instead of picking one fixed action for a step, you can give it a goal in plain language, an explicit list of which tools it's allowed to use, a maximum number of steps it can take to get there, and a dollar budget it can't exceed. The agent works toward the goal within those bounds — it's not an open-ended prompt with unrestricted reach into your workflow's capabilities.
Wired to real capabilities, not one action per service
An agent step can be configured against any of the platform's AI capabilities — dozens of them, routed through the same AI gateway every other AI feature on the platform uses — rather than being locked to a single bound action per integration the way most competitor "AI nodes" are.
Why the bounds matter
Give an agent only the access it needs, and cap what it can spend before it starts, is standard human-in-the-loop practice for anyone running AI actions unattended — and it's exactly what tools[] and the dollar budget field enforce directly, not as a suggestion in the documentation. A misbehaving or looping agent step can't quietly run past a step count or a spending cap; both are hard limits on the step itself.
If you've wanted an AI step that can actually take multiple actions toward a goal — not just transform text once — this is that step, with the guardrails built in rather than left to you to add.