Build an ACT automation with your agent
You say what should happen. Your agent writes the ACT automation. There is no editor to learn and no code to write. The agent writes the code, tests it, and switches it on once you confirm. To change it later, you tell the agent again.
For a job that needs the agent's judgement every run, such as a weekly summary, you do not need any of this: ask the agent for an agent automation.
Who do I ask?
- Boosty. Tell Boosty in Chat, and it hands the job to the agent that fits it.
- An agent directly. Open the agent's chat, for example the Sales agent for lead work.
- From Automations. Open the ACT automations tab and click New automation. The chat opens with "When a record is created, run code that" in the message box. Finish it and send it. With one agent picked in the scope, the chat is with that agent, otherwise with Boosty.
The agent needs Automations set to Allowed in Connections and security.
What do I tell the agent?
Three things, in your own words:
| Say | For example |
|---|---|
| When it runs | "Every weekday at 7:00 Prague time", or "When a lead is created" |
| What it works with | "the Leads module", or "our Stripe account" |
| What it does with the result | "Post the hot ones to our Sales channel in Slack." |
All together: "When a lead is created, score it from 1 to 10 by company size and budget, and post the leads above 7 to our Sales channel in Slack."
The more exact you are, the fewer questions the agent asks. Name the module and the fields, the threshold, the channel and the time zone. When a module name fits more than one module or space, the agent asks which one you mean before it builds.
What does the agent do then?
- It reads your data model. It looks up the modules and fields it works with, so the code uses their real names.
- It checks the other service's documentation. For a call to another app, it writes the request from that service's own documentation. If it cannot find it, it asks you for a link.
- It asks for keys it needs. If the automation talks to another service, the agent sends you a one-time link to save the key as a secret. The key never appears in the chat.
- It tests the code. Each test runs in an isolated environment. It first tests on sample data, without reaching any other service. When a test fails, the agent fixes the code and tests again.
- It asks for your yes. Before it reads from another service for the first time, it tells you which service it signs in to and what it reads. A test never writes anything there. Before it switches on a schedule or a webhook, it shows you the exact schedule or webhook.
- It switches it on. The automation appears in Automations, with a one-sentence summary of what it does, and the agent tells you in the chat.
How do I check that it works?
Ask the agent to try it on real data before it goes live: "Run it once on the last five leads and show me the result." After that, every run is listed with when it ran and what it returned. See Runs.
Admins can open the automation's page and read its Code too. You never have to.
How do I change it?
Open the automation and click Edit with Boosty for an ACT automation, or Edit with <agent> for an agent automation. The chat opens with "Change the ACT automation … so that" in the message box. Finish the sentence: "it runs at 8:00 instead", or "it also includes the lead's country." The agent changes it, tests it and asks you again.
When the last run failed, the button is Fix with Boosty (or Fix with <agent>), and the message asks the agent to find out why and fix it.
An agent changes only the ACT automations it built itself.
How do I pause or remove it?
- Pause. Click Pause on the automation's page. An ACT automation asks you to confirm. It does not run until you click Resume, and scheduled runs are skipped meanwhile.
- Remove. Ask in Chat: "Delete the automation Score new leads."
Can I build automations from Claude or ChatGPT?
Yes. Connect your AI tool with Connect AI (MCP) and ask it the same way. It builds, tests and asks you to confirm exactly like an agent in Chat. An admin signs in with OAuth, or uses a token with All permissions.
Next: Triggers and webhooks.
