AI agents
An AI agent is a helper you set up by describing a job in plain words. It works only on the sites, data and tools you allow, and it stops to ask you before anything big.
One job, start to finish
The agent does the work. You approve what matters.
What it is
An AI agent is a helper that works for your business. You describe a job in plain words, like "answer new enquiries from my contact form", and the agent does it using the AI model you pick.
You decide what it can reach, when it starts and which actions must wait for your approval. An agent can start on a schedule, from a webhook, from an incoming email address, from an account event, or when you ask it to. Each time it runs, you can read what it did, step by step.
An automation (a Flow, a Journey or a Function) follows steps you set up in advance. An agent gets a job in words and works out the steps itself, inside the limits you set. Use an automation when every run should be identical, and an agent when the job needs reading, deciding or writing. The two work together: a Flow can include a step that calls an agent. See Automations.
What you can use it for
Each of these uses features that exist today. You can set one up by describing it to the chat, or by filling in the form.
When someone sends a form on your site, the agent reads the entry, adds the person to your CRM and drafts a reply. Sending the reply waits for your approval.
Set up with: the Event (when something happens) trigger "A visitor submitted a form on one of your sites", access to the site and to your CRM's data, and Can send email.
Every Monday the agent reads your table of new leads or orders and writes a short summary. You read it in the agent's History. If you also switch on Can send email, it can email the summary, and that email asks for your approval unless you untick it.
Set up with: the On a schedule trigger (preset Every Monday at 9:00) and Bases they may work on.
The agent looks at new rows in a table, fills in missing details and flags the ones that need you, for example anyone who asked about pricing.
Set up with: Bases they may work on and a trigger of your choice, or run it yourself from the Run tab.
Give the agent its own email address. Anything sent or forwarded to that address becomes a task for it, such as "summarise this supplier email and list what I need to reply to".
Set up with: the Incoming mail (email address) trigger and its Add address button.
Add Telegram under Channels, choose which agent to Chat with:, and message your bot. The agent does the task and replies in the chat.
Set up with: Channels > Add Telegram.
Let the agent work in apps you have connected, for example reading or writing in a tool you use every day. You tick which connected apps it may use, and you can allow only some of their actions.
Set up with: Authorised connections on the agent (connect the app first under Automations > Connectors).
How it works
This is what you do in Serenities, under Automations > AI Agents.
01Describe the job
Open Automations. It opens on AI Agents, with the chat on the left. The chat asks "What should your AI agent do?" Describe the job in plain words and it builds the agent for you, including which tools it may use and when it runs.
Under the chat box are three buttons: Build an agent, Schedule it and What did they do? They fill in a starter sentence for you.
Prefer to click? Press Build Specialist AI Agent and fill in the form yourself. Step 2 shows it.
02Choose what it may touch
Pick the sites and bases the agent may work on. For each site, choose Read-only or Can edit. New sites start as Read-only. Then switch on only the extras the job needs: Can send email, Can fetch the web and Can run your apps and sites. When you build an agent with the form, all three start off. If you ask the chat to build it, the chat chooses them for the job you describe.
To let it use an app you have connected (for example Gmail), tick the app under Authorised connections. You can expand an app to allow only some of its actions.
You can see all of this on the agent's Permissions tab, and change it any time with Edit on the agent's page.
03Choose what needs your approval
Under When it wants to do something big, choose Ask me first (recommended). Then tick the kinds of action that must wait: Publishing sites, Sending emails & campaigns, Payments and Deleting things. When the agent reaches one, it stops and shows you exactly what it wants to do. Nothing happens until you decide.
Payments always wait, whatever you choose. If you pick Let it decide on its own, the agent does the other big actions without asking, so use it only for jobs you trust.
You can approve or decline in the dashboard, from the email we send you, or in Telegram if you add it under Channels. Whichever you answer first counts. The link in the email works for 72 hours. Add a note under Tell it why (optional) when you decline, and the agent skips that action and wraps up.
04Teach it skills
A skill is a short guide for the agent: when to use it, and the steps in plain words. On the agent's Skills tab, press Write one, or Import skill to bring in a SKILL.md or .zip file, or ask the chat to write it. When a run goes well, press Save as skill in its History to keep what the agent did.
The agent also saves short facts to its Memory as it works, up to 50. You can Edit or Forget any of them. If you do not want it to learn, switch off Learns as it works when you set it up.
05Choose when it runs
On the agent's Triggers tab, pick what starts it:
Or run it yourself: type a task on its Run tab and press Run, or type "@" and its name in any chat. You can also ask your own Claude or ChatGPT app to run an agent, once you have added Serenities there as a connector (see Use your own AI). Switch the agent off with the Active switch and its triggers stop starting it.
06Check its work
Each run shows a short Report, a checklist under What it did, how long it took and the credits it used. Press Details to see every step, Stop to end a run that is going, or Ask AI about this run to ask the chat about it. Every run stays in History.
You also set limits for each agent: Spend limit per run (credits), Runs per day (max), Actions per run and Time limit per run. When a run reaches a limit it stops and says so.
Limits
No. You describe the job in the chat, or fill in a short form. Skills are written in plain words.
An automation (a Flow, Journey or Function) runs steps you set up in advance, the same way each time. An agent is given a job in words and works out the steps within the limits you set. They work together: a Flow can include a step that calls an agent. See the Automations page.
You get an email whenever an agent needs your approval or has a question. Add Telegram or more email addresses under Channels and they get the request too. If a run that started on its own fails, you are told the same way, and every run sends its report to your channels. You will also see an amber Needs your approval badge on the agent in your list.
Runs use your AI credits, and each run shows what it used. You set a spend limit per run and a limit on runs per day. If you add your own AI model key, runs on that model use no credits. See the Pricing page for what each plan includes.
Yes. Connect the tool once under Automations > Connectors, then tick it under Authorised connections on the agent. You can allow only some of its actions. Using a connected app asks for your approval while the agent is set to Ask me first.
You can build agents on any plan, up to 10 on Free. They run on paid plans.
Anything set to ask first only happens after you approve it, so check the card before you press Approve. An email that has been sent cannot be taken back. If you do not want an agent to change a site, allow it as Read-only, and it cannot.
