Sistava

Taskade AI Agents: How to Build Your First One

How-to — — by Mahmoud Zalt

A practical walkthrough of building your first custom AI agent in Taskade, from the Agents tab to the system prompt, knowledge, tools, and commands.

Most people arrive at Taskade's agent builder after watching a demo where somebody types one sentence and a working assistant appears. Then they open the real thing, see four tabs and a dozen fields, and stall on the first one.

This walkthrough covers the parts that actually matter for a first agent, in the order you meet them. It also flags the fields you can safely leave alone until you know what the agent is for.

The important idea before you start: in Taskade, an agent is a saved configuration, not a running process. You are writing down who it is, what it knows, and what it is allowed to touch. Nothing happens until you chat with it, call one of its commands, or wire it into an automation.

Where do AI agents live in Taskade?

Agents live in the Agents tab of a workspace, alongside your projects, templates, and automations. That placement matters for a practical reason: an agent belongs to the workspace it was created in, so a teammate in a different workspace will not see it.

If you work across several workspaces or client spaces, decide up front where the agent should live. Taskade documents a copy function that duplicates an agent into another workspace, with an option to carry its knowledge across, but it is less annoying to create it in the right place first.

What are the ways to create an agent?

Taskade documents four starting points, and they suit different moods. Pick the one that matches how clearly you can already describe the job.

For a first agent, open a template even if you do not intend to keep it. Reading somebody else's finished configuration teaches you more in two minutes than the empty form will in twenty.

Build your first agent

  1. Open the Agents tab and create a new agent — Start from the AI Agent Generator, a template, or from scratch. All three land you in the same editor afterwards.
  2. Fill in the General tab — Name, Description, Persona, Tone, Introduction, and Language. The Description is the closest thing to the system prompt, so this is the field worth your attention.
  3. Choose the AI Model — Taskade offers thinking modes including Auto, Standard, Thinking, Reasoning, and Advanced, backed by models from OpenAI, Anthropic, and Google. Auto is a reasonable default until you have a reason to change it.
  4. Turn on Knowledge and add sources — Enable the Knowledge toggle, then add files, workspace projects, external links, or YouTube videos. Skip this if the agent does not need to answer from your material.
  5. Add tools if the agent needs to do things — On the Tools tab, use Add tool to give the agent abilities such as creating projects and tasks or calling automations. An agent with no tools can still talk, it just cannot act.
  6. Create commands for anything you repeat — On the Commands tab, give the command a Name and a Prompt, and pick a Command Mode. Default runs the prompt, Plan and Execute works through structured goals.
  7. Test it in chat before you trust it — Open the agent's chat and run the questions you actually expect. Taskade provides a Regenerate control so you can get a fresh answer without retyping your message.

The Description field carries almost all the weight. Taskade's own guidance for the quick setup is to write a prompt that describes the agent's job, and that is genuinely the whole trick.

What should go in the description and persona fields?

Write the job, the boundaries, and the output shape. Three short paragraphs beat one long one, because you will be editing this field repeatedly and you need to find the line you want to change.

The job is one sentence: what this agent handles and for whom. The boundaries are what it must never do, such as inventing a price or promising a delivery date. The output shape is the format you want back every time, such as three bullets and a suggested next step.

Persona and Tone are separate fields for a reason. Persona is who the agent is, Tone is how it sounds. Keeping them apart means you can make an agent friendlier without accidentally changing what it believes its job is.

If the setup feels like a lot, there is another shape

Filling four tabs, choosing a model, wiring tools, then testing and re-testing your wording is real work, and it is work you own forever. That is a fair trade when you want precise control over a specialist assistant, and Taskade gives you a lot of control for the effort.

That set-up work is exactly what we removed at Sistava. Rather than configuring a persona, a model, and a tool list, you hire an AI Employee and tell it in plain English what you want handled. The tools are connected already, it asks you when it is unsure, and it sets itself up.

How do you actually use the agent once it exists?

An agent is not only a chat window. Taskade documents several entry points, and the ones inside your projects are where the time savings show up.

If you only ever chat with your agent, you have built a slightly personalised chatbot. The commands and the bulk-run path are what turn it into something that saves an afternoon.

Limits and gotchas the docs do not lead with

Two things surprise people early. The first is the plan boundary, the second is how much iteration a good agent takes.

The second surprise is honest and worth saying plainly: independent reviewers who have built agents in Taskade consistently note that agents need tuning before they are reliable. That is not a criticism of Taskade, it is true of every configurable agent builder. Budget a few rounds of editing the description and re-testing.

One more practical note. Because the agent belongs to a workspace, an agent that answers from client A's documents should not be the same agent you point at client B. Separate agents per client is tedious to set up and much safer to live with.

FAQ

How do I create an AI agent in Taskade?

Open the Agents tab in your workspace and create a new agent. Taskade documents four routes: the AI Agent Generator, their EVE assistant, a pre-built template, or starting from scratch. You then fill in the General tab (name, description, persona, tone, model), optionally enable Knowledge and add sources, add tools on the Tools tab, and create slash commands on the Commands tab.

Which AI model should I pick for a Taskade agent?

Taskade exposes thinking modes including Auto, Standard, Thinking, Reasoning, and Advanced, backed by models from OpenAI, Anthropic, and Google. Auto is a sensible starting point. Move up to a reasoning mode only when you can point at a specific answer that came back shallow, since heavier modes use more of your credit allowance.

Can I use Taskade AI agents on the free plan?

Taskade's pricing page on 17 August 2026 listed the Free plan as including 1 agent along with a one-time credit allowance, with unlimited agents on the paid tiers. Because these numbers change, check the current pricing page before you build a workflow that depends on a specific count.

What is the difference between an agent command and a normal prompt?

A command is a saved prompt with a name, which you call with a forward slash inside a project or a chat. It also has a Command Mode: Default simply runs the prompt, while Plan and Execute works through structured goals. Commands are how you stop retyping the same instruction.

Can other people use an agent I built?

Yes. Taskade documents shared agent chats where you invite people into a conversation with the agent, and publishing an agent with a public link, including appearance options and password protection. Agents can also be copied into another workspace, with an option to bring their knowledge along.

Why does my agent ignore what I told it?

Usually the instruction is buried or the agent has no source for the answer. Split a long description into short paragraphs so each rule is visible, add explicit negative rules such as what to do when it does not know, and confirm the relevant document is actually attached under Knowledge rather than assumed.

Sources

Build the agent if you want the control. If what you actually wanted was the work done, that is a different product category, and it is the one we build.