On 6 October 2026, Zapier moved the capabilities that lived in standalone Agents into a single AI by Zapier step inside the normal Zap editor. An agent used to be a prompt, a list of tools, a trigger, and permission to act, sitting outside filters, paths, and Zap History. Now those pieces live inside a workflow and run on the same task system as the rest of the account.
For a small team already paying for Zapier, the agent is no longer a side app. It becomes another step you can control.
Model choice is new. Standalone Agents did not let you pick. AI by Zapier offers Standard, Advanced, and Premium tiers, or a named model. Approval is also more useful now. Instead of an all-or-nothing permission model, Require approval before running is set per tool. A step can read a sheet automatically, for example, and stop before it sends a message to a client.
Cost is measured in Zapier tasks rather than a separate activity meter. Default models are free. Other tiers bill 1, 3, or 5 tasks per step, and the usage is visible in Zap History. If one run crosses 75 tasks, Zapier pauses it for a human instead of letting it continue indefinitely. That cap is a useful guardrail against runaway agent loops, although Zapier says it will become adjustable later.
Migration is a rebuild, not a switch. Eligible agents show a banner linking to Zapier's migration guide. Most prompts and tools carry over, but knowledge sources are not fully supported yet. HubSpot, Asana, Zendesk, Zoom, and Jira are currently missing from the AI step. On-demand chat becomes a webhook trigger, while one agent calling another maps to a Sub-Zap.
Take a three-person studio that drafts a client update from a form and then emails it. Previously, the agent could own the entire job, including the send. Now the AI step can choose the status and write the subject into structured fields. A filter can drop empty rows. A normal email action can send the template. Approval can sit only on the email tool. Zap History then shows the model tier, the tools called, and the tasks consumed next to every other Zap step.
Zapier's pitch is that you get reasoning where it helps and predictable workflow logic everywhere else. That is a fair description for this kind of job, but it does not mean the AI step is cheaper on every run. A Premium step billed at 5 tasks and called twice will burn through a plan much faster than a filter. The 75-task pause is a safety rail, not a budget limit.
The useful part is straightforward: the editor you already know, per-tool approval, and a bill you can inspect in Zap History. The harder cases are agents that depend on HubSpot, Asana, Zendesk, Zoom, or Jira as knowledge sources. Move one agent that can send or publish this week. Leave the rest until Zapier says they are eligible.
