​

Overview

Building an agent in WorkLLM is simpler than it sounds. You describe the task in plain words, and WorkLLM helps with the rest. This article walks you through each step, from your first agent to turning it on and running it. It is useful whether you are building an agent yourself or running one that a teammate built. No technical background is needed. Every section tells you what to do, what you should see, and what to watch out for. If you are new to agents, please start with Agents overview. Here is what we cover:

• What to set up before you start.
• How to create your first agent, with a quick example.
• How to write a good prompt.
• How to review and edit the steps, including the model and the app for each step.
• How agents use knowledge and memory.
• How to start an agent on a schedule, or when something happens in an app.
• How approvals work for actions that write data.
• How to test an agent and run it.

​

Before you start

You do not have to prepare everything before you begin. Nothing here is compulsory. Think of this as a short checklist, not a set of rules.

• Connect the apps your agent will use. Agents work with tools like Google Workspace, Slack, Jira, HubSpot, Notion, Salesforce, and many others through MCP. It helps to connect the ones you plan to use, but you can also connect them later if you change your mind. So if you have, say, only Outlook connected right now, that is fine. You can add more whenever you need them. Each connection costs 10 credits, and that is a one-time charge per app.

• Decide which task to automate first. Think about the work that takes most of your week, such as weekly reports, follow-up emails, or inbox sorting, and pick one that repeats in the same way each time. That makes the best first agent. 

​

Create your first agent

Let us build something simple together. This walkthrough creates a morning industry digest. The agent only reads and summarizes, so nothing goes out to a client. It is a safe first agent, and it doubles as the example for the rest of this article. To create the agent:

1. Describe the task. Start a new agent and write what you want in plain language, the way you would explain it to a new colleague. You can type or paste this example:

Every weekday at 8:00, search the web for the latest articles about the logistics industry. Pick the five most useful, write two sentences on each, and email the list to our engagement team. 

1. Answer a few questions. If something is missing or unclear, WorkLLM asks before it builds anything. For this agent, it might ask who should get the email, or whether you prefer certain publications. Just answer in plain words, the way you would to a person. 
2. Review the steps. WorkLLM puts together a step-by-step plan. For this agent, expect steps like these: search the web for new articles, pick the five most useful and summarize each, and email the list. Every step shows the app it uses, the knowledge it draws on, and its instructions. Change anything that does not match what you meant.
   
3. Save the agent and test it. Save the plan as an agent. Run it once yourself and check the result (see “Testing an agent before turning on automation”). When you are happy with it, turn on the 8:00 weekday schedule.

What you should see: a saved agent, and an email with five article summaries in the inbox you chose. Email Draft:  One small thing to know: web search adds 1 to 3 credits per search, on top of the normal run cost. For the full picture, see How WorkLLM credits work.

​

Writing a good prompt for agent creation

You do not need a perfect prompt for creating an agent. WorkLLM asks questions when something is missing before creating an agent, so you cannot really get it wrong. That said, the more of these you include, the fewer questions you will have to answer, and the closer the first plan will be to what you imagined:

• What the agent does, and who it is for. Say the job and the audience.
• What starts it. A time, an event in an app, or someone running it by hand.
• Where the information comes from. Name the app, folder, channel, or document.
• What the result looks like, and where it goes. For example, a five-slide deck emailed to the account lead.
• What it must not do. For example, “send the draft to me, and do not email the client.”

Seeing this side by side usually helps. Here are a few weak prompts next to stronger versions of the same idea. A few tips that save time later:

• Give each agent one job. Several small agents are easier to test and reuse than one large one.
• Use your own names for things, like the folder, channel, or client list.
• Add a past report or proposal to the agent’s knowledge, so it follows the format your team already uses.

If the plan is not what you expected, just edit the prompt or the steps, and review the plan again.

​

Reviewing and editing the steps

After your answers, WorkLLM builds the plan. Nothing runs until you have looked it over, and you can edit any step. To review and edit the plan:

1. Open the plan and read it from the first step to the last.
2. Select a step to open it. Change its instructions, app, knowledge, or model.
3. If the plan does not match how your team works, reorder, add, or remove steps.
4. Save your changes.

When you read through the steps, here is what to check:

​

Choosing a model for each step

Every step can use its own model, so a simple step does not need an expensive one. WorkLLM gives you more than 200 models to choose from. To choose a model:

1. Open the step.
2. Open the model picker. It shows a cost level for each model, from one dollar sign to four.
3. Select a model, or leave the Recommended setting on, and WorkLLM picks one for you based on cost and quality.

Not sure which to pick? We recommend leaving the model setting on Recommended. WorkLLM will then choose the most suitable model for each step, balancing cost and quality for you. If you still prefer to choose a model yourself, here is our guidance. Tips:

• If you are not sure which model to pick, we recommend leaving the model setting on Recommended. 
• After a test run, check the credits used, and swap lighter models into the steps doing simple work.
• A model that runs on your own API key (BYOK) does not use WorkLLM credits for the model itself, and you can set a BYOK model on any step.

​

Attaching an integration to each step

Each step that works in an app shows which app it uses. A step can either read data, or act on it, for example by posting a message or updating a record. To attach or change an app on a step:

1. Make sure the app is connected (see “Before you start”).
2. Open the step.
3. Select the app the step should use.

Here are a few examples of what that looks like: Agents can act in more than 100 natively connected tools, and in almost any other tool through MCP. Just contact the team and they will help set it up. Steps that send or update things are covered in “Approvals for actions that write data.”

​

Adding knowledge to an agent

You decide what knowledge each step uses. An agent does not pull knowledge from anywhere on its own : whatever you add to a specific step is exactly what that step uses. You can add knowledge to a step in three ways:

• Organization memory, for company-wide material like brand guidelines or policies.
• A file you attach, in any standard format: PDF, DOCX, spreadsheet, image, or any other standard file format.
• A link.

To add knowledge to a step:

1. Open the agent and select the step.
2. Click the + icon to see the knowledge options : Organization memory, a file, or a link
3. Choose Organization memory, a file, or a link.
4. Check that the knowledge now shows on the step.Add only what the step needs. A proposal step might use your proposal template, while a reporting step might use last month’s report as a format example.

Adding organization memory: Added organization successfully in the step.

​

Triggers

An agent can start in three ways: on a schedule, when something happens in a connected app, or when someone runs it. In fact, almost any action in a connected app can be the trigger. An agent needs something to set it off, and you have three options:

• A time you choose. For example, every weekday at 8:00.
• An event in a connected app. For example, a new email arrives or a file is saved.
• A person. Someone opens the agent and runs it by hand.

Once you have set one of these, the agent waits for the moment to arrive, then starts on its own. Almost anything that happens in a connected app can work as the trigger. Manual Trigger edit: Trigger Edit by AI: Here are some common triggers and what the agent does with them: Adding or changing a trigger
You can set a trigger yourself or ask the AI to do it. To set it manually:

1. Open the agent and go to its trigger setting.
2. Choose a schedule for a time based trigger, then set the days and time.
3. Choose an app event for an event based trigger, then pick the app and the event, such as a new email arriving.
4. Save, then run a test.

To change it through AI, tell the agent what you want in plain words, for example “Run this every weekday at 9:00 instead of every Monday” or “Start this when a new email arrives from a client.”

​

Approvals for actions that write data

Reading data carries very little risk. Writing data is where a second look matters. Examples are sending an email, posting a message, updating a CRM record, or creating a file. For each connected app, you decide whether its actions run automatically or wait for your approval first. Approval is set for each app Approval applies to the whole app. It does not depend on the agent, the event, or the use case.

• If you choose to always approve an app’s actions, every agent that uses that app runs those actions without a pause. For example, with email set to always approve, every email from any agent goes out right away. This applies whether the email goes to you, a teammate, or a client.
• If you keep manual approval on, you approve each action from that app. This applies to every agent that uses the app.

Approval cannot be set for a single agent. Please keep this in mind when you choose a setting.

​

A sensible starting point

Because the setting applies to the whole app, choose it based on the riskiest thing any of your agents might do with that app.

​

Staying safe with client communication

If you are not ready for an agent to contact clients directly, you can tell the agent to send its results to you first. You review them and forward them only when you are happy. This works no matter which approval setting you use. Once you trust the agent, you can change its instructions to send directly.

​

Testing an agent before turning on automation

Run the agent yourself a few times before you switch on a schedule or trigger. This is the safest way to catch anything that is off. To test an agent:

1. Run it once with a real but low-stakes input, like a recent report or a sample client.
2. Check each step’s result against the plan. Did each step do what you described?
3. Check the output. Compare the format, tone, and numbers with your last manual version.
4. Send test results to yourself. Keep test runs away from clients and shared channels. 
5. Fix what needs fixing. Edit the step or the prompt, and run it again.
6. Check the credits used (see Agent reports and credit usage). If the cost is higher than you expected, switch lighter models into the simple steps. 
7. Turn on the schedule or trigger.

Testing uses credits, the same as any run.

​

Running an agent

You can run an agent in two ways: When the user asks, or automatically.

• When the user asks: open the agent, fill in any details it asks for, and run it. Anyone with permission to run the agent can do this.
• Automatically: turn on its schedule or trigger. After that, the agent runs in the background, and nobody has to remember to start it.

Here is how the two compare: Each run uses credits. To see the result and cost of every run, see Agent reports and credit usage.

​

Frequently asked questions

• Do I need to write a perfect prompt for agent creation?
  No. WorkLLM asks clarifying questions before it builds the plan, and you review and edit every step before anything runs.
• Can I edit a step after the agent is built?
  Yes. You can edit steps before an agent runs, and you can edit the agent later. Changes apply to every future run. For the full steps, see Editing and deleting agents. 
• Can each step use a different model?
  Yes. You can choose a lighter model for simple steps or a stronger one for analysis, and you can set a BYOK model on any step.
• What can start an agent?
  There are three ways an agent can get started:
  a. A schedule you set. You tell the agent when to run, like “every Monday at 9:00”, and it starts by itself at that time. No one has to click anything.
  b. Something happening in a connected app. For example, a new email arrives, a file is saved to a folder, or someone submits a form. The moment that happens, the agent kicks in and starts working.
  c. A person running it. You or a teammate opens the agent and clicks run, whenever you need it.

In short: an agent starts because the clock says so, because something happened in one of your apps, or because a person asked it to.

• Can an agent send emails or update records on its own?
  Yes, in the apps you connect. You choose whether it runs on its own or holds for your approval.
• How do I test an agent without sending anything to real people?
  Run it manually and send the results to yourself. Check the output, then turn on automation
• My agent did not run when I expected. What should I check?
  Check that its schedule or trigger is turned on, that your workspace still has credits, that the app it uses is still connected, and the agent’s run history. You can find the run history in Agent reports and credit usage. 
• The result is not what I expected. What can I do? Open the plan and read each step against what you meant. Tighten the instructions, name your sources, and add a past example to the agent’s knowledge. Then test it again.