What are Zapier Agents and how do they differ from Zaps?
Zapier's help centre describes agents as AI-powered assistants that do work on your behalf, independently or when you interact with them. A classic Zap follows the exact steps you set; an agent receives instructions, a set of tools and optional knowledge, and decides which tools to call and in what order.
That flexibility is the point and the risk. An agent can handle an enquiry that does not fit a neat rule, but it can also take an action you did not expect. Reviewing agents is therefore a core skill, not an afterthought. If your process is predictable, a normal Zap, covered in the Zapier automation tutorial linked below, is usually the safer choice.
What is happening to Zapier Agents?
Zapier is migrating Agents from agents.zapier.com to AI by Zapier, a step built into the Zap editor that brings tool calling and reasoning into a single AI step you can combine with normal triggers, filters and branching. Its migration guide says migration converts each agent into a Zap with a native trigger and one AI by Zapier step holding your prompt, instructions and tools.
Zapier states that it has not set a shutdown date for Agents and will email affected users with notice. Existing agents keep running until you turn them off, so if you migrate, turn the old agent off after publishing the new Zap, because both running at once can cause duplicate actions. Zapier lists a few gaps that are still rolling out, including some knowledge source apps and a direct equivalent for chat-triggered agents.
How do Zapier Agents use tools and data sources?
An agent uses tools, which are app actions you give it, and knowledge sources, which are data it can reference. In Zapier Agents you add tools from the Instructions field with Insert tools, choosing a default tool or an app and one of its actions, and you add knowledge sources for data the agent should look things up in rather than packing it into long instructions.
Each tool is a permission. An agent with 'send email' can send email; an agent with 'update deal' can change your sales records. Give the smallest set of tools the job needs, prefer read actions to write actions and keep knowledge sources to material the agent genuinely needs.
What does a well-scoped agent look like?
Take a fictional 20-person bookkeeping firm, Cobaltline, whose clients email questions about deadlines. The goal: when a client emails, find their next filing deadline in the firm's deadlines spreadsheet and draft a reply for the account manager.
Badly scoped: tools to read the inbox, send email, edit the spreadsheet and update the CRM, with the instruction 'handle client emails'. The agent could reply directly to clients, change a deadline or edit a record, and nobody would know until a client complained.
Well scoped: tools to read the spreadsheet and create an email draft only, a knowledge source containing the firm's standard wording, and instructions that say 'never send; always create a draft; if the client is not in the sheet, say so in the draft and stop'. A person sends every reply. In AI by Zapier you can go further by marking any tool that changes data with Require approval before running.
How do you review what a Zapier agent did?
Review an agent by reading its run history: Zapier's activity pages show each run's status, the apps used and when it ran, and the run details show what information the agent used and how it interpreted and carried out your instructions.
Find the activity
Open All activity for every agent, a pod's activity for a group, or Activity on a single agent's page. You can pin important runs.
Read the reasoning
Expand a run to see how the agent interpreted your instructions. Look for steps where it guessed.
Check the outcomes
Open the draft, record or message it created and compare it with what you would have done.
Keep evidence before deleting
Zapier says deleting an agent deletes all its activity, so export or note anything you need first.
In AI by Zapier, use Zap history
After migration, runs appear in Zap history with the tools called, model tier and tasks used.
Can you make a Zapier agent ask for approval before it acts?
Yes, in AI by Zapier. Zapier's migration guide describes a per-tool setting, Require approval before running, that pauses the run before that tool executes; tools run without pausing by default. Runs also pause for review when they exceed a task limit, which defaults to 75 and can be raised. Admins can require Zaps with AI steps to pass an approval process before they are published, and can turn tool calling off at account level.
Zapier's announcement describes approval in the standalone Agents product as all-or-nothing, which is one reason to plan your review around the new step. For high-stakes actions, you can also add a Human in the Loop Request Approval step in the surrounding Zap.
How does Tosheo.ai teach Zapier Agents?
Tosheo.ai covers Zapier Agents in the AI agents development course and the AI automation course, starting with when an agent is the right tool and when a plain Zap is better. Lessons use fictional firms like Cobaltline, and the hands-on task asks you to scope tools, write instructions with stop conditions and review a run. The rubric marks least-privilege tools, approval points and the quality of your review notes. Lessons take 10 to 15 minutes, and every plan starts with a 7-day free trial with no credit card needed.
Sources and scope
Zapier: Zapier Agents product page, checked 19 October 2026 ↗
Zapier Help Center: review your agent's activity, checked 19 October 2026 ↗
Zapier blog: Zapier Agents is now AI by Zapier, checked 19 October 2026 ↗
Zapier Help Center: Human in the Loop Request Approval, checked 19 October 2026 ↗