Gumloop Tutorial: Build Your First AI Agent Without Code

👁 115 reads

Gumloop is a platform for building AI agents and automations from building blocks, without writing code, and with an assistant that can build a flow for you based on a description. This tutorial takes you from your first flow to automation that runs on its own, following Gumloop’s official documentation.

Before you start: describe the task on paper

The best automations start with a description of work that someone does manually today. Write down step by step what a person does: where they get data from, what they do with it, where they save the result, and how often. This description will be useful both when building manually and when working with the Gummie assistant, to whom you can simply paste it.

Also consider which steps really require AI. Copying data or filtering rows can be done by regular nodes, without credit costs, and you save AI for where you need to understand text.

Step 1. Two basic concepts

Gumloop documentation bases everything on two concepts. Agents are AI assistants that use tools to solve tasks. Flows let you run agents and actions automatically: on a schedule, in bulk, or after an event. It’s worth keeping this distinction in mind from the start, because it determines where you’ll begin.

Step 2. Quickest start: Gummie assistant

Gumloop has a built-in assistant called Gummie that creates a flow based on a description of what you want to automate. You can also ask it what tasks are worth automating in your role, then open a ready-made template. For people who don’t know where to start, this is the shortest path.

Step 3. First flow manually

If you prefer to understand how it works, build your first flow yourself. On the canvas you add an input node, for example a text field with a website address. Nodes are building blocks, and connected in sequence they create a flow. Most flows have a similar pattern: fetch data, prepare it for AI, process through AI, and return the result, for example to a spreadsheet.

After adding a node, you run the flow and check whether the data looks as it should. Test after each new node, not just at the end.

Step 4. Connecting to applications

Gumloop connects with popular applications, for example Google Sheets, Slack, or Salesforce. You connect accounts by logging into the given service. Over half of basic nodes, for example text operations, filters, and some integrations, don’t consume credits, so you mainly pay for AI steps.

Step 5. Automatic execution

Gumloop has three ways to run flows without your involvement:

• triggers, which react to an event in a connected application, for example a new row in a spreadsheet,
• interfaces, meaning a form or page overlaid on the flow, through which others provide data without seeing the automation,
• schedule, for example daily at eight o’clock.

Step 6. Credits under control

The best place to check credit usage according to the documentation is the statistics panel in organization settings. In a conversation with an agent, the current cost is visible next to its name. The documentation recommends starting with recommended model settings, and turning on the most powerful models only for tasks requiring deep reasoning.

Example: research companies from a list

You have a spreadsheet with names and websites of twenty potential clients. You build a flow: a node reading the spreadsheet, a node fetching the page content of each company, an AI node summarizing what the company does and whether it fits your offer, finally saving the result in a new column. You test on three rows, then run it on the whole set.

After such a test, it usually turns out that AI handles summarization well, but assessing fit to your offer requires a more precise instruction. So you add to the AI node what your company does and what clients you’re looking for, ideally with two examples of well-matched and poorly-matched companies. After this correction, the results are clearly more accurate, and you can run the flow for subsequent lists without changes.

Most common mistakes

• Running on an entire spreadsheet without testing on a few rows.
• Most expensive model for a simple task.
• Automation without checking AI results in the first days.

How to evaluate your first automation

For the first week, compare automation results with what a person would do. Take ten cases and check them manually. If AI makes mistakes in one or two, correct the instruction in the node or add an example of a correct result. If it makes mistakes more often, the task may require better prepared input data.

Also check credit usage in the statistics panel after a few days. This lets you estimate monthly cost before you run the flow at full scale.

When the flow works, save it as a starting point for the next ones. Similar tasks, for example researching another list or summaries from another source, you’ll then build by copying and making small changes, instead of starting from a blank canvas.

// read moreGumloop: No-Code AI Agents - How It Works & Pro Plan CostAI ToolsGumloop Use Cases: Lead Research, Ticket Sorting & ReportsGuidesAI Agents: What They Are, How to Build One & Where They FailGuides