Skip to content
Product Agents

Search Product Agents

Start with a guide, or search a topic

    New here? Read the five guides in order

    Subscribe

    Zapier Agents & AI by Zapier

    Zapier is turning its agent into one step inside a Zap.

    Zapier is folding its standalone agents into its automation editor, where an agent is one AI step billed in tasks. Here is what that step can do, how its approvals work, and where runs go wrong.

    Zapier Agents showing a Meeting prep agent: a Schedule by Zapier every-day trigger, step-by-step instructions that use Google Calendar: Find Event, Google search and Slack: Send Direct Message, an expected outcome, and an Agent preview panel with a Test agent button.
    Image: Zapier · Agents product page

    At a glance#

    What it does

    An AI by Zapier step inside a Zap reads a plain-language prompt and calls tools in a loop until the job is done. Tools can be any of 9,000+ app actions, web search, and up to 20 knowledge sources. The older standalone Zapier Agents product is being converted into these Zaps.

    Price

    Billed in Zapier tasks. One run costs the model tier’s rate for the step plus the same rate per tool call. Tiers: Standard 1x, Advanced 3x, Premium 5x, your own API key 1x. Professional starts at $19.99 a month billed yearly for 750 tasks; Team at $69 for 2,000. The standalone Agents add-on is billed in “activities” instead: 400 a month free, or 1,500 for about $33.33 a month.

    Model

    Chosen by tier. Standard: GPT-5 nano and mini, Gemini 2.5 Flash. Advanced: GPT-4.1, Claude 4.5 Haiku, Gemini 2.5 Pro. Premium: GPT-5 and GPT-5.4, Claude Sonnet 4.5 and 4.6, and Claude Opus 4.5 through 4.8. Enterprise can route through its own AWS Bedrock keys. The standalone Agents page describes “a fine-tuned model for research and task management” without naming it.

    Takes actions?

    Yes, through any connected app action. A per-tool “Require approval before running” switch pauses the run before that tool and shows the filled-in fields; it ships off. A run that reaches 75 tasks (adjustable to 500) pauses for review.

    Claimed results

    Zapier’s governance page cites Okta automating 13% of support escalations and Remote automating 2,200+ days of work a month. Its Agents page says 2.2 million+ companies use Zapier and one customer’s agent produced over 2,000 leads in a month.

    Launched

    Zapier Central became Zapier Agents a year after launch. May 22, 2025: agents shifted from chat to automation. April 23, 2026: Agents generally available with Enterprise MCP. June 15, 2026: AI by Zapier moved to model-tier pricing. September 23, 2026: Next Gen Zaps early access. The Agents-to-AI by Zapier migration guide was updated October 1, 2026.

    Owner / Status

    Zapier, private; co-founder and CEO Wade Foster. Standalone Agents still run at agents.zapier.com with no shutoff date set. Zapier says it will email users before one.

    Why it matters#

    Zapier sold a standalone agent product, first as Zapier Central and then as Zapier Agents. It had its own site, its own billing unit and its own approval advice. In 2026 Zapier began converting every one of those agents into a Zap: a trigger, then one AI by Zapier step that calls tools in a loop.

    Zapier is turning its agents into one step inside a Zap. The step is billed in tasks like any other. A tool inside it can be switched to ask a person first. A run that reaches 75 tasks pauses. The admin rules that say which apps and actions any Zap may touch now apply to the agent too.

    How it works#

    What happens when a request reaches Zapier Agents & AI by Zapier.

    1. Trigger

      Schedule by Zapier fires every day at 8:00 AM. After migration, the Meeting prep agent from Zapier’s product video is a Zap with this trigger and a single AI by Zapier step holding its four instructions.

    2. Reason and call tools

      The step reads the instructions and works through them. Google Calendar Find Event gets today’s meetings, a Google search runs for each outside attendee, then Slack Send Direct Message sends the summary. Every call happens inside the one step.

    3. Pause, if a switch is on

      With “Require approval before running” on for the Slack tool, the step shows Delayed and the reviewer gets an email and an Activity notice. They see the message as filled in and click Approve or Reject. Reject stops the run. The switch is off unless someone turned it on.

    4. Hit the brake

      If the run reaches 75 tasks, the Zap pauses and asks a person before continuing. An admin can raise the cap to 500.

    5. Finish and bill

      The Slack message sends. Zap history lists the tasks used, broken down by tool call, model tier and step. On the Advanced tier, for example, one step plus three tool calls is 12 tasks.

    6. Turn off the original

      Until the old agent is switched off, both run. Zapier’s own guide says to turn off the agent after publishing the Zap, or the message sends twice.

    The decisions worth studying#

    What they chose, why, and what it costs.

    1It dissolved the standalone agent into one step in the editor

    Zapier Agents lived at agents.zapier.com, apart from the Zap editor. Zapier’s reason for ending that, in its own words: “you could not combine your agents with the full library of Zap triggers, actions, filters, branching logic, and automation history.” So Zapier converts each agent automatically. The result is a Zap with a native trigger and a single AI by Zapier step holding the prompt and the tools.

    My read: this is the right call for a company whose asset is 9,000 app connections and a run history. The agent gains everything a Zap has, including the governance controls below.

    The costs are specific. Tool calls stay inside the one step, so a Zap that looks like two boxes may be making a dozen writes. Agents you started by chatting to have no equivalent trigger; Zapier says to use a webhook. Knowledge sources like HubSpot, Zendesk and Jira aren’t supported until Q4 2026. Until you turn the old agent off, both run. And there is no shutoff date.

    2It bills the agent in tasks, by model tier, with a 75-task brake

    Since June 15, 2026, an AI by Zapier run costs its tier’s rate for the step plus the same rate for each tool call. The rates: 1x on Standard, 3x on Advanced, 5x on Premium, or 1x with your own API key. Standard can’t use tools. Zapier’s formula is (1 × model rate) + (tool calls × model rate). For example, a Premium step that calls four tools is 25 tasks, against 750 in a starter Professional month.

    My read: once tool calls loop, the model decides at run time how much a run costs. Zapier added a brake for that: if a run reaches 75 tasks, the Zap pauses and asks a person; admins can raise it to 500.

    Two things to check. Zapier’s June 15 notice says new steps default to Advanced; its model-tier pricing page says Premium. And full-Zap tests currently consume tasks like production runs. Zapier says a fix is coming.

    3Approval is a switch on the tool, off by default

    Zapier has given agents three ways to ask a person. The oldest, for standalone agents, is a sentence in the instructions: ask for my confirmation through Slack “before continuing.” The second is a separate Human in the Loop Request Approval step. It can be set to continue when the reviewer declines or never answers. Set that way, Zapier’s status docs file approved, declined and timed-out runs all under Success. Only the step’s Decision field says “approved.”

    The third is the one to copy. Each tool in an AI by Zapier step has a “Require approval before running” switch. On, the run pauses before that tool “regardless of how specific your prompt is,” shows the fields, and a Reject stops it. Zapier recommends it for tools that create, update or delete.

    It ships off. In Zapier’s own product video, a Meeting prep agent runs daily and sends a Slack message at step 4, with nothing in between that waits for a person.

    Zapier Agents · instructionsProduct screenshot
    A Meeting prep agent’s trigger, “Schedule by Zapier: Every Day”, and instructions with an Insert tools button: “1. Access your calendar for today's scheduled events.” (Google Calendar: Find Event), “2. Identify all meetings with external participants from the retrieved events.”, “3. For each external participant, search for relevant information (e.g., company, role, recent interactions)” (Google search), “4. Send a direct message to you on Slack containing the compiled information” (Slack: Send Direct Message). Expected Outcome: a Slack message “every morning at 8:00 AM” listing external participants and what was found about them.
    A schedule starts the run. Four steps later Slack sends. No switch between them is on. Zapier Agents product page video · 0:21

    4Admins govern the agent the way they govern every Zap

    On April 23, 2026, Zapier made Agents generally available alongside a set of Enterprise controls. App Access Controls allow or block apps across a workspace at the API level. Action Restrictions can “allow reads from Salesforce while blocking creates or updates.” Managed App Connections are owned by IT. Bring Your Own Model runs through AWS Bedrock, and logs stream to Datadog, Splunk or a SIEM. Wade Foster’s line that day: “new ideas are already outrunning the guardrails around them.”

    My read: this is the payoff of making the agent a step. One policy layer covers Zaps, agents, MCP clients like Claude and ChatGPT, and SDK apps, because they all run the same actions.

    The limits: every one of those controls is Enterprise only. On Professional and Team, the per-tool switch and the task cap are what you have. And a standalone agent’s record dies with it: delete the agent and its activity is deleted too.

    What to watch#

    • Two defaults in the docs

      Zapier’s June 15 notice says a new AI step defaults to Advanced (3x). Its model-tier pricing page says Premium (5x). Open the step and check before you estimate a month.

    • No shutoff date, and two copies running

      Standalone Agents keep running until Zapier sets a date, and a migrated agent runs beside its original until someone turns the original off. Count what each one sent.

    • Next Gen Zaps add an agent that fixes your Zaps

      Announced September 23, 2026 and in early access on paid plans. An agent “watches your runs, diagnoses failures, and can apply or suggest a fix.” Decide who approves a fix to a workflow that writes to customers.

    • Success is a status, and approval is a field

      Set to continue, a Request Approval step reports Success whether the reviewer approved, declined or never answered. Any later step that should wait for a yes has to read the Decision field.

    What it means for your team#

    What to decide before you ship an agent like Zapier Agents & AI by Zapier.

    • Platform

      Put the approval before the write: switch on “Require approval before running” for each tool that creates, updates or deletes. Branch on the Decision field, since a step’s status only says that it ran.

    • Finance

      Budget in tasks: the tier rate for the step plus the same rate per tool call, capped at 75 a run unless raised. Check the tier each step is set to, and remember that tests consume tasks today.

    • Security

      The app, action and connection controls that make an agent governable are Enterprise only. Below that tier, decide which tools an agent may hold at all, because the switch is your only gate.

    • Support

      Copy Zapier’s connection errors, which say whether access expired, was revoked, lacks permission or was unshared. Its “Needs action” status mixes those with clarifications and approvals, so each needs a different owner.

    Steal this

    1. Pause the run before the one tool that writes, and show the filled-in fields.

    2. Branch on the reviewer’s decision. A finished step proves only that it finished.

    3. Cap tool calls per run with a hard number, and pause for a person at the cap.

    Reference#

    The evidence behind this teardown: recordings, sessions, what Zapier Agents & AI by Zapier gets right and wrong, and every source.

    Walkthrough · 2 frames

    From a daily schedule to the Test agent button#

    Two crops of one frame from the video on Zapier’s Agents product page, captured September 28, 2026. A vendor recording, not a live test.

    1. Step 1 of 2: The builder sets a schedule and steps

      A Meeting prep agent’s trigger, “Schedule by Zapier: Every Day”, and instructions with an Insert tools button: “1. Access your calendar for today's scheduled events.” (Google Calendar: Find Event), “2. Identify all meetings with external participants from the retrieved events.”, “3. For each external participant, search for relevant information (e.g., company, role, recent interactions)” (Google search), “4. Send a direct message to you on Slack containing the compiled information” (Slack: Send Direct Message). Expected Outcome: a Slack message “every morning at 8:00 AM” listing external participants and what was found about them.
      Image: Zapier Agents product page video · 0:21

      1. A schedule, not a person, starts each run: “Schedule by Zapier: Every Day”.

      2. The builder writes plain-language steps with tools inline: Google Calendar: Find Event, then Google search.

      3. No step in the instructions asks a person before step 4 sends the Slack message.

    2. Step 2 of 2: Agent preview offers a test run

      The agent’s header with an Activity tab, app icons, a toggle and Share, above the Agent preview panel: “Test this agent”, “Test your agent and chat with your agent to make sure it’s running properly.”, and a Test agent button.
      Image: Zapier Agents product page video · 0:21

      1. Activity holds each run’s details and tool calls; deleting a standalone agent deletes that activity too.

      2. Test agent runs the agent by hand in the preview, which lists the tools the agent used.

    Sources#

    12. ZapConnect 2026zapier.com

    Sources checked September 28, 2026.

    Next teardownLinear Agent

    Linear lets a person hand an issue to an AI agent while a person stays the assignee, so someone is always responsible for the result.