Build your first workflow: draft, approve, post to Slack¶

This page builds a small workflow on the canvas, using only the forms. The agent drafts a message, waits for you to approve it, then posts it to a Slack channel. It takes about ten minutes, most of which is the one-time Slack setup.

By the end you will have used the three things you need for most workflows: an action step (the model writes something), an approval step (a human decides), and a tool step (something happens in the real world).

Before you start¶

You need the canvas running and a model key. If you started the Docker stack with make up, the canvas is at http://localhost:4310 and it opened the starter agent from workspace/agent.yaml. If you use kampong dev, see Getting started.

You also need a Slack bot token and a channel. The setup is in Demo: support triage: create an app, add the chat:write scope, install it, invite the bot to a channel, and note the channel ID.

Put the values in .env:

OPENROUTER_API_KEY=...
SLACK_BOT_TOKEN=xoxb-...

A spec never contains a secret, only the name of the variable. You can also set the values from the canvas instead of editing .env: click Secrets in the status bar, then Set next to SLACK_BOT_TOKEN and paste the token. It is saved on your machine, takes effect immediately, and is never shown again.

The starting point¶

The starter agent has one step, greet, that asks the model to answer the input. On the canvas you see a trigger, the agent, and one workflow step, with the YAML next to them. Everything you do below changes that YAML, and you can read it at any point.

1. Add the Slack tool¶

A tool describes something the agent can do. It does nothing until a workflow step uses it.

1. Open the ⋮ menu in the top bar and click Add tool (or + Tool under the step palette).
2. Set Tool kind to Slack.
3. Name: post_to_slack.
4. Token (env reference): ${SLACK_BOT_TOKEN}. This is the variable name, not the token.
5. Channel: the channel ID, for example C0123456789, or a name like #support-triage for a public channel. A private channel needs the ID.
6. Text: {{ greet.text }}. This inserts the output of the greet step.
7. Save.

A tool node appears on the canvas and a tools: entry appears in the YAML.

2. Add an approval step¶

1. Choose the step kind in the palette on the left, or open the ⋮ menu and click Add step.
2. Set Step kind to Approval.
3. Step ID: review.
4. Message: Post this to Slack? {{ greet.text }}.
5. Save.

The run will stop at this step and show the message to whoever is running it. If they reject, the run ends and nothing is posted.

3. Add the tool step¶

1. Choose the step kind in the palette on the left, or open the ⋮ menu and click Add step.
2. Set Step kind to Tool.
3. Step ID: send.
4. Tool name: post_to_slack.
5. Save.

Steps run in the order they appear, so the workflow is now greet, then review, then send.

4. Check your setup¶

Click Checks in the status bar, turn on Also check credentials, and run the checks. The panel lists each variable the spec needs by name, and it asks Slack whether the token is accepted. It never displays a value. Fix anything marked as failing before you run.

5. Run it¶

1. Click Run in the top bar.
2. In Input, type something for the model to answer, for example Write a two-line welcome for new support customers.
3. Click Run.

The trace shows each step as it finishes. At review an Approval required dialog shows the drafted text. Click Approve and the send step posts it. Open the channel in Slack to see the message.

What you built¶

This is the whole workflow in YAML. You can open it in the </> drawer (read only) and edit it in any editor, and the canvas updates.

version: "1.0"
agent:
  id: hello_agent
  name: "Hello Agent"
  role: "Friendly assistant"
  goal: "Greet the user warmly and answer one question in a concise, helpful tone."
  model:
    provider: openrouter
    name: liquid/lfm-2.5-2.6b:free
    api_key: ${OPENROUTER_API_KEY}
  tools:
    - name: post_to_slack
      action: slack_post_message
      token: ${SLACK_BOT_TOKEN}
      channel: "#support-triage"
      text: "{{ greet.text }}"
  workflow:
    - step: greet
      action: generate_text
      inputs: [input]
    - step: review
      type: approval
      message: "Post this to Slack? {{ greet.text }}"
    - step: send
      type: tool
      tool: post_to_slack

Going further¶

• Give a step its own prompt. On an action step, the optional How it should behave, Model and Creativity (temperature) fields override the agent's role and goal, model and sampling temperature for that step only. They need version: "1.1" at the top of the spec.
• Branch on the result. Add a Condition step to take a different path, for example to escalate instead of sending. See Workflows and conditions.
• Run it from the terminal or as a webhook service. See Demo: support triage, which does this with a fuller spec.

If something goes wrong¶