Every capability that made Zapier Agents powerful—tool calling, reasoning, autonomous action—now lives inside a single AI step in the Zap editor, using AI by Zapier. Here, I'll map each Zapier Agents feature to its equivalent in AI by Zapier, so if you're coming from Zapier Agents, you know exactly where to click.
What's changed between Zapier Agents and AI by Zapier
Zapier Agents consisted of four things: a prompt, a set of tools, a trigger, and permission to act. That bundle worked, but it sat outside the platform, so your agent couldn't use, for example, Zapier's triggers, filters, paths, or run history.
The same four ingredients now live in a step you drop into any Zap, running on the infrastructure that already executes millions of tasks a day and lets you access the rest of the Zapier platform.
Moving from Zapier Agents to AI by Zapier
What you built in Agents | Where it lives now |
|---|---|
Agent instructions | The prompt field on the AI by Zapier step |
Tools the agent could use | The tools picker inside the AI by Zapier step—add as many as the job needs |
Agent trigger (schedule, webhook) | The Zap trigger—any of 9,000+ apps, not just a handful |
Chat / on-demand agent | |
One agent calling another | Sub-Zap by Zapier as the trigger on the child Zap |
Granular Human Intervention | Require approval before running, set per tool. (New: Agents were all-or-nothing; now you set approval granularity per action) |
Model choice | Standard, Advanced, and Premium tiers—or name a specific model. (New: Agents didn't let you pick; now you control cost and capability per step) |
Agent run history | Zap History, alongside every other automation you run |
Agent cost | Tasks, on the same task-based pricing as the rest of Zapier; no more guessing what an activity is—we now have a universal task billing system. |
Knowledge sources | The knowledge sources your AI step can reference while reasoning (not yet supported in AI by Zapier: HubSpot, Asana, Zendesk, Zoom, and Jira) |
5 things you can do with AI by Zapier that you couldn't do with Zapier Agents
AI by Zapier gives you all sorts of combinations that were impossible when agents lived on their own island.
Mix agentic and deterministic steps
Let the AI decide which record to update, then use a normal Zapier action to send the confirmation email exactly the way it should read every time. That means you get reasoning where it adds value and predictability everywhere else. This allows you to have better control of your workflow and reduce costs over time.
Read more: How Zapier can help you with valuemaxxing, not tokenmaxxing
Gate risk per tool, not per agent
Reading data from a spreadsheet doesn't carry the same weight as messaging a client. With Require approval before running toggles, you control the granularity—letting your agent read data to reason autonomously, while forcing a human check before taking action that can't be undone.
Get structured output you can actually map
Define the fields you want back—things like a category, a sentiment score, or a subject line—and they arrive as clean, mappable outputs. You won't need to parse a paragraph of prose in a downstream step.
Remove the guesswork
Let AI automatically detect when you might need to add tools for the agent and add them to your step. This removes the guesswork of what's needed for your agent to operate and takes the work off your plate. You can always manually add or edit yourself if the output isn't quite right.
See what happened
Every run lands in your Zap History with the tools that were called, the inputs and outputs, the model tier used, and the tasks consumed. It's the same interface you already use for the rest of your Zaps.
Understand your AI costs before you pay
With AI by Zapier, AI steps bill as tasks. You can use default models for free, and other model tiers carry different multipliers, so an AI step can cost 1 task, 3 tasks, or 5 tasks. You can see the breakdown per run in your Zap History rather than discovering it on an invoice.
You can also cap it. Set a limit on tasks per run, and if a single run crosses 75 tasks (this will be adjustable in future), it pauses for a human instead of looping quietly. We know runaway agents were a concern with Zapier Agents, and this is our solution.
Already using Zapier Agents? Migrate to AI by Zapier
You don't have to rebuild your agents from scratch or from memory. The migration is to recreate your agent as an AI by Zapier step: it's the same goal and same tools, now living inside a Zap where you can wrap deterministic guardrails around it. Most of what you had carries over—and you gain the control and observability the standalone product couldn't offer.
A few capabilities are still rolling out. Where something isn't available yet, there's usually a clean workaround (for example, you can mimic an on-demand run today by triggering the agent from a webhook).
If your Zapier Agent is eligible for migration, you'll see a banner at the top of the agent that will lead you to step-by-step guide into the new product. See this help article to learn more.
