How to Migrate Zapier Agents to AI by Zapier: A Practical Checklist

Introduction

Zapier Agents no longer live in their own product. They now live inside Zaps as a single AI by Zapier step: good news for anyone who feared a rebuild, and a trap for anyone who clicks through migration without checking what changed. The converted Zap works, but its billing and approval behavior differ from the old agent. This tutorial covers the migration as it actually works: convert one agent, audit its prompt and tools, gate dangerous actions, tame the task cost, and only then publish.

Quick answer

Go to agents.zapier.com, click Migrate your Zap, review the AI step prompt and tools, turn on approval for write actions, preview, test, then publish. Migration converts agents automatically; it does not split tools or enable approval by default.

Evidence

Zapier’s own migration guide states the core promise plainly: “Zapier automatically converts your agents to Zaps. You do not need to rebuild anything from scratch.” Each converted agent becomes a Zap with a native trigger step and a single AI by Zapier step, and that AI step contains the original prompt, tool calls, and AI reasoning, all in one place. That makes this a checklist, not a rebuild, and it is why review matters: anything you do not check in that one step ships as-is.

The platform behind the step matters too. Zapier’s announcement maps the old agent pieces onto the new model: the trigger is now any Zap trigger, drawn from more than 9,000 apps, rather than the handful the standalone Agents product offered. A chat or on-demand agent becomes a webhook trigger, one agent calling another becomes a Sub-Zap by Zapier trigger on the child Zap, and run history moves to Zap History next to every other automation.

Billing is where most migrations quietly go wrong. Since June 15, 2026, AI by Zapier steps have been priced by model tier rather than as a flat action: Standard at 1x, Advanced at 3x, and Premium at 5x, with each tool call billed at the same tier rate. Tasks per run equal one tier-rate unit for the step itself, plus one per tool call. An Advanced step that calls two tools therefore costs 9 tasks per run, and the same step on Premium costs 15. A single run that reaches 75 tasks pauses for human approval instead of looping on, which caps a runaway agent but not your monthly total. The tier shown on your step is what you will be billed, not the tier you remember from June.

Three primary sources anchor this tutorial. Zapier’s official Migrating from Agents to AI by Zapier guide documents the conversion flow, the review and publish sequence, and the Enterprise note. Zapier’s announcement post Zapier Agents is now AI by Zapier documents the mapping from old agent pieces to new Zap pieces, including per-tool approval and the model tier choice. Zapier’s guide to adding tools to your AI by Zapier step documents the tools picker used below. All three pages were verified as live in October 2026.

Prerequisites

• A Zapier account with agents listed at agents.zapier.com.
• Editor access to the Zaps migration creates, plus permission to publish them.
• A list of every agent, including what each one is allowed to change: which apps it writes to, which messages it sends, and which records it deletes.
• A low-risk agent to migrate first, ideally one that only reads data or sends an internal notification.
• Access to Zap History, so you can compare task counts before and after migration.

Step-by-step guide

Step 1: Inventory your agents before you touch anything

Open agents.zapier.com and list every agent. For each one, write down its trigger, its tools, its knowledge sources, and the most dangerous action it can take. Knowledge sources matter because Zapier’s announcement notes that some sources, including HubSpot, Asana, Zendesk, Zoom, and Jira, are not yet supported as knowledge sources in AI by Zapier, so an agent that leans on one needs a workaround first. The most dangerous action tells you where approval must go in Step 4.

Step 2: Migrate your lowest-risk agent first

Each agent shows a Migrate your Zap link. Click it for your lowest-risk agent only. A dialog box appears once your Zap is ready to review. Click Open Zap to enter the Zap editor. You now have a Zap with a native trigger step and a single AI by Zapier step. Do not publish yet: the review in the next steps is what makes it safe, and one completed migration gives you a realistic task count and approval pattern to copy to the rest.

Step 3: Review the AI step prompt, instructions, and tools

In the Zap editor, click your AI by Zapier step to review your prompt and tools in the Configure tab. Read the prompt as if you had never seen the agent before, and check that the instructions still match what the business wants now that the agent sits inside a larger Zap. Check the tools list against your Step 1 inventory: every tool the agent had should be present, and none you do not recognize. If a tool is missing or a knowledge source is unsupported, fix that gap before testing. Click Edit to make changes, and keep the prompt tight: vague instructions produce extra tool calls, billed at the tier rate.

Step 4: Turn on per-tool approval for every dangerous action

This is the step migration does not do for you. Approval is set per tool, not for the whole step, and the Require approval before running toggle is off by default for each tool. A migrated tool that sends email, charges a payment, updates a CRM record, or deletes a row will therefore run without pausing unless you turn approval on for that specific tool. Turn approval on for every tool that creates, updates, or deletes data, and for anything customer-facing or irreversible. Leave it off for read-only tools, such as looking up a record. For a full walkthrough, see our guide to setting up per-tool approval in a Zapier AI step. The Zap is not safe to publish until every write tool has an explicit approval decision.

Step 5: Split deterministic actions into native steps where it pays

Migration does not split your tool calls into separate native Zap steps: your tools continue to run inside the AI step, the same way they did in your agent. That is fine for judgment calls, where the AI decides which record to update or what to write. It is wasteful for fixed actions, where the outcome is the same every time. If your agent always sends the same confirmation email after it finishes, move that send into a native email step after the AI step. The native step bills as one ordinary task and runs with exactly the same wording every time, instead of consuming a tier-multiplied tool call. Keep reasoning inside the AI step, move predictable output outside it, but do not split genuinely judgment-based actions.

Step 6: Preview the AI step with sample data

In the Zap editor, click Preview to run the AI step with the sample data. Review the AI’s output, which tools it called and what each one did, and how many tasks the run consumed. If the preview calls a tool you expected to stay silent, your prompt is looser than you thought, and tightening it now is cheaper than paying for the extra tool calls on every live run. If the task count surprises you, work the formula backwards: one tier-rate unit for the step, plus one per tool call, at the tier shown on your step.

Step 7: Test the whole Zap end to end, then publish

Click the Test tab, then click Test Zap to run the complete workflow end to end. The test runs every step in order, including native steps you added in Step 5, so it catches ordering problems a step preview cannot: a native step that expects an output field the AI step does not produce, or a trigger mapping that changed during conversion. Review the results in the test panel. If anything looks off, adjust the prompt or tool settings directly in the step and test again. Once all steps pass, click Publish, watch the Zap in production, and only then migrate the next agent.

Step 8: Watch Zap History for two weeks and tune the tier

After publishing, check Zap History for the migrated Zap regularly for the first two weeks. For each run, note the tools called, the task count, and any run that paused for approval. Compare the task count against the old agent’s cost pattern, because the tier multiplier means the same behavior can cost three or five times what the step count suggests. If a Zap uses tools for work a native step could do, split that work out using Step 5. If a Zap runs on Premium but only needs classification or extraction, drop the tier to Standard, which costs 1 task and does not support tool calls, or to Advanced if it still needs tools. If runs keep pausing at the 75-task limit, tighten the prompt, add a stop condition, or split the work across steps.

Troubleshooting

The migrated Zap behaves differently from the old agent

Start with the tools list, not the prompt. A tool that did not carry over, an unsupported knowledge source, or a changed trigger mapping will all look like prompt problems but are not. Compare the AI step tools against your Step 1 inventory and the trigger against the original agent trigger. If a knowledge source is on Zapier’s not-yet-supported list, replace it with a native lookup step before the AI step, or with a supported source, and test again.

A migrated tool runs without asking for approval

This is the default, not a bug. The Require approval before running toggle is off by default for each tool, so a tool runs without pausing unless you turned approval on for that specific tool. Open the AI step and turn approval on for every write, send, update, and delete tool, one tool at a time. Do not assume approval on one tool covers the others, or that the old agent’s all-or-nothing setting carried over, because the new model is per-tool by design. Confirm a gated tool actually pauses a test run.

The task bill jumps after migration

Work the formula first: one tier-rate unit for the step plus one per tool call, at Standard 1x, Advanced 3x, or Premium 5x. An Advanced step with two tool calls costs 9 tasks per run, and at a thousand runs a month that is 9,000 tasks from one Zap. Check the tier shown on the step, because defaults have changed during 2026 and that tier is what you are billed. Then count tool calls per run in Zap History. If the count is high, tighten the prompt, split fixed actions into native steps, or drop the tier.

A run keeps pausing at the task limit

A run that reaches 75 tasks pauses for approval instead of continuing, which is the safeguard working as designed. Treat repeated pauses as a symptom. The usual causes are a prompt without a stop condition, a tool that returns data the agent keeps re-checking, or a loop between two tools. Tighten the prompt with an explicit stop condition, split the looping portion into a native step with a fixed result, or reduce the tools available to that step. Raising the limit without fixing the loop just moves the bill.

The migrated Zap fails on a step the old agent handled

Check the mapping between the AI step’s outputs and the steps after it. A native step added after migration, or a trigger field that changed during conversion, can expect an output field the AI step does not produce in its new form. Open the test panel from Step 7, find the failing step, and compare the field it expects with the field the AI step actually returned. If output fields need to be explicit, define them in the step configuration so downstream steps receive clean values instead of free text. For longer workflows that need to survive failures and resume cleanly, our guide to Zapier durable workflows covers retry and resume patterns for migrated agents. Our guide to migrating away from Zapier Functions before the cutoff covers that separate deprecation.

FAQ

Do I have to rebuild my Zapier Agents from scratch?

No. Zapier converts each agent to a Zap with a native trigger step and a single AI by Zapier step, keeping your original prompt, instructions, tool calls, and AI reasoning. Migration is a review, approval, and cost-checking job.

Does migration split my agent’s tools into native steps?

No. Tools keep running inside the AI step exactly as they did in the agent. Splitting a fixed action, such as sending a standard confirmation email, into a native step is optional, and worth doing because a native step bills as one ordinary task instead of a tier-multiplied tool call.

How much does a migrated step cost?

It depends on the model tier and the tool calls. Standard bills at 1x, Advanced at 3x, and Premium at 5x, with each tool call billed at the same rate. An Advanced step with two tool calls costs 9 tasks per run. Check the tier on your step before publishing, and Zap History after publishing.

Is per-tool approval on by default?

No. The Require approval before running toggle is off by default for each tool. Turn it on for every tool that creates, updates, deletes, sends, or charges, and leave it off for read-only tools. A migrated Zap with ungated write tools is not ready to publish.

What happens to chat agents and agents that called other agents?

A chat or on-demand agent becomes a Zap with a webhook trigger, and one agent calling another becomes a Sub-Zap by Zapier trigger on the child Zap. Run history moves to Zap History, so you can inspect migrated runs alongside your other automations.

What about Enterprise agents?

If your Enterprise plan includes Agents, Zapier’s guide says to contact your account representative for your migration plan. Confirm the plan before publishing migrated Zaps widely.

Conclusion

Migrating Zapier Agents to AI by Zapier is not a rebuild, but it is not a one-click finish either. Convert one agent, review its prompt and tools, gate every dangerous action per tool, split the fixed work into native steps, and let two weeks of Zap History tell you whether the tier and the tool count are right. Do that, and the migrated Zap is cheaper, safer, and easier to inspect than the agent it replaced. For the broader picture, The Complete n8n Guide shows how agent workflows fit into a larger stack. Start with your lowest-risk agent today.

Frequently asked questions