How to Build and Deploy Next Gen Zaps Using AI Agents, MCP, and Natural Language

A practical field guide from Automation Ace.

The short answer

To build a Next Gen Zaps workflow, connect an AI agent to Zapier (through Zapier's MCP server in a client like Claude, ChatGPT, Cursor, Claude Code, or Codex, or with the built-in agent in the Zapier editor), then describe the workflow in plain language: the trigger, the apps, and what each step should do. The agent picks app connections with you, writes the workflow as code, saves a draft, tests it with sample input, and publishes a version when you approve.

Not sure whether you need Next Gen Zaps? Start with classic Zaps vs Next Gen Zaps.

Before you start

• Access: Next Gen Zaps are in early access by request on paid plans.
• An agent: an MCP-capable client connected to Zapier, or the Zapier editor's agent. New to MCP? See API vs CLI vs MCP vs SDK.
• App connections for every app involved. See connecting apps to Zapier.
• A written spec, even a short one. Agents build what you describe, so clarity matters.

Step 1: describe the workflow

Include four things:

1. Trigger: a schedule, an app event, or on-demand only.
2. Apps and accounts: which Gmail, which CRM, which Slack workspace.
3. Steps and rules: what happens to each item, including branches, loops, and exceptions.
4. Failure behavior: what to do when a step fails or data is missing.

"Every weekday at 7am, get all open deals in HubSpot that have had no
activity for 14+ days. For each deal, look up the owner, and if the deal
amount is over $10,000 post a message to #sales-alerts; otherwise email
the owner a summary. If a lookup fails, log it to the 'Stale Deal Errors'
Zapier Table and keep going. Send me a count of each outcome at the end."

Name edge cases up front, such as blank fields, duplicates, or time zones. See handling blank values.

Step 2: choose connections and let the agent draft

The agent walks you through selecting the right app connections, then writes the workflow as TypeScript and saves it as a draft. You can keep more than one draft at a time, which is useful for trying two approaches. The editor shows a read-only diagram of the logic.

Step 3: test with sample input

• Ask the agent to run the draft with realistic sample data, including edge cases.
• Review each step's output, not just whether the run succeeded. See why a run can succeed while nothing happens.
• Ask the agent to explain the logic back to you in plain language and compare it with your spec.
• Test failure paths deliberately. See testing without live data.

Step 4: publish and turn it on

When the draft behaves correctly, publish it. Each publish creates a version, so you can review changes over time and start a new draft from an earlier version if you need to roll back. Then enable the workflow and watch the first live runs closely. See version control for no-code automations.

Step 5: set up Agentic Management

Agentic Management lets an agent review each run, diagnose failures, and prepare a fix. Decide how much autonomy it gets: many teams start with “diagnose and propose” and require a person to approve any change to production. Pair it with your own monitoring and error handling.

Best practices for AI-built workflows

• Least privilege: give the agent and workflows only the connections they need. See the security checklist.
• Protect sensitive data: use AI Guardrails and avoid pasting secrets into prompts.
• Keep humans on risky actions: build pause-for-approval steps before payments, deletions, or customer messages.
• Document intent: save the prompt and a plain-language summary with each version. See automation documentation.
• Iterate in drafts, never by editing a live version directly.
• Pilot before migrating classic Zaps; run both in parallel and compare outputs.

For what this model unlocks, see 7 things Next Gen Zaps can do that classic Zaps can't. Request access to Next Gen Zaps and build your first workflow from a clear spec.

Frequently asked questions

How do I build a Next Gen Zap?

Connect an AI agent to Zapier through an MCP client or use the Zapier editor's built-in agent, describe the trigger, apps, and steps in plain language, choose app connections, let the agent draft and test the workflow, then publish a version and turn it on.

Which AI tools can build Next Gen Zaps?

Zapier describes building from MCP-enabled AI tools such as Claude, ChatGPT, and Cursor, coding agents like Claude Code and Codex, or the built-in agent in the Zapier editor.

What should I include in a Next Gen Zap prompt?

The trigger, the apps and accounts involved, what each step should do including branches and loops, edge cases like blank or duplicate data, and what to do when something fails.

How do I test a Next Gen Zap before deploying?

Save it as a draft and have the agent run it with realistic sample input, including edge cases and failure paths. Review each step's output and ask the agent to explain the logic back to you.

What is Agentic Management in Next Gen Zaps?

It lets an agent watch a published workflow, review each run, diagnose failures, and prepare fixes. You decide how much it can do on its own.