Agents

An agent is a standing instruction for one site, such as "every Monday, summarise last week's visitors and the new comments". On the schedule you set, Modulify starts an AI agent with the instruction, the agent reads the site with read-only tools, and its report is emailed to you. Scheduled jobs call one of your URLs. An agent reads the site itself and tells you what it found.

Over MCP and the API, agents are called agent automations, after their tools.

Where agents live

Click More in the tab strip, then Agents. The address ends in /agents. The tab has two lists:

• Agents shows each agent with its switch, its schedule, when it runs next, how its last run went and whose it is. The three dots menu on a row holds Run now, Run history, Change with chat and Delete. An agent that is off carries an Off tag, and the reason it switched off shows under it.
• Runs shows every run, newest first, with its status (Running, Done, Failed or Skipped), when it started, and Scheduled or Run by with the name of whoever started it by hand. Open a run to read its Report by the agent, its error, What it read (each tool the agent called, with the ones that failed in red), whether the report was emailed, and the steps and credits it used. Run history on an agent shows only its runs, and the chip above the list shows them all again.

When a site has at least one agent, a bar under the chat tabs says how many there are and what they are doing, such as 2 agents · 1 running or 1 agent · next run in 3h. Click it to open the tab.

In the Modulify app for iPhone and Android, Agents is in More too, with the same two lists, see On the mobile app.

Agents are for people who can edit the site. On a site where your role is Content editor or Viewer, Agents is disabled with Your role on this site does not include this., and so are the buttons on agent cards in the chat.

Create one

Ask for one in the editor chat, or click Create with chat in the Agents tab, which starts your message with Set up an agent that . The chat shows a New agent card with the name, the instruction, the schedule and the next few runs, and nothing is saved until you agree in the chat. Once you do, an Agent created card replaces it.

A site can hold 3 agents.

From an AI client connected over MCP, use create_agent_automation. It returns the same preview first and creates nothing until it is called again with confirm set to true.

Change one

Only the person who set an agent up can change its name, instruction or schedule, because its reports go to them. For anyone else, Change with chat is disabled and reads Only the person who set it up can change it.

Change with chat starts your message with Change the "<name>" agent so that . The chat shows an Agent changes card with the agent as it would be, and nothing changes until you agree. A new schedule moves the next run, and a run already going keeps the old instruction. Over MCP, use update_agent_automation.

Anyone who can edit the site can switch an agent on or off, run it now or delete it.

Schedule

The schedule takes the same options as scheduled jobs: every hour, every day, every week, every month, or a cron expression, in the timezone you choose. Runs cannot be closer together than 5 minutes, and a site can start at most 10 runs a day across all its agents. A day is a UTC calendar day, and a run started by hand counts toward it too.

If a scheduled time is missed, for example while the site is locked, that run is skipped and the next one goes ahead on time.

What a run does

Each run starts with your instruction, the site's name and the time of the run in the schedule's timezone. The agent can use read-only tools for this one site's analytics, logs, comments, configuration and data. It cannot change the site, publish it, or send email to anyone but you. Tools whose names mention a secret, key, token, password, credential or URL are not offered to it, so it never gets a stored secret or a link that opens a file without a login.

A run may use up to 12 steps. After 10 minutes it stops using tools and writes its report from what it has gathered. The report keeps up to 8,000 characters.

The report is emailed to the person who set the agent up, from Modulify, with the subject Agent report: <agent> on <site>. The email says it was written by your agent. The report is written by an AI model from the site's data, so check important numbers before you act on them.

An agent never writes in a chat and never contacts your visitors. The report is all it sends.

Credits

Every run that does work is charged to the workspace's AI credits, at the same rate per credit as the editor chat. That includes a run that fails after it starts, for the work it did before it failed. Each report email is charged at the site email cost per send. Tool calls are not charged on their own, because what a tool returns is read by the model and so is already part of the AI cost. A run that is skipped, or that fails before it starts, costs nothing. Agent runs show in Usage as Agent run, under AI.

Before each step, the run checks that the workspace can pay for that step, the work the run has already done and the report email. If it cannot, the run stops with The workspace ran out of credits during this run, so the run stopped. It does not count as a failed run. A run therefore normally stops before it would owe more than the workspace holds.

A charge is recorded once, even when it is retried after an interruption.

Emails about an agent

Besides its reports, the person who set an agent up is emailed when a run fails, when a scheduled run is skipped, when a run is interrupted, and when the agent switches itself off after failing. Each of these sends at most one email in 24 hours, and none of them is charged.

No email is sent when an agent switches off because its site was deleted or because the person who set it up left.

Run now

Run now starts a run outside the schedule. It is refused while the agent is off, while a run of it is still going, while the site is locked, when the workspace has used its credits, and when the site has used its runs for the day. Each run counts toward the daily limit above.

When it switches off

An agent switches itself off after 3 failed runs in a row, and the reason shows on it. A successful run resets the count, and switching it back on clears the reason and the count.

These do not count as failed runs, and they do not count toward the daily limit either:

• A run interrupted on Modulify's side, for example by a restart. It says The run was interrupted on Modulify's side before it finished, so there is no report this time. The next scheduled run goes ahead, and Run now runs it again.
• A run the AI model did not answer.
• A run that ran out of credits.

A run that has not finished after two hours counts as failed.

An agent also switches off when its site is deleted, and when the person who set it up leaves the workspace or no longer has an active account. It cannot be switched back on after that, so set up a new one.

Locked sites

On a site that is locked for going over its plan's site limit, runs are skipped and the agent's owner gets an email. Creating, changing, switching on and Run now are refused. Switching off and deleting still work.

If Modulify switches an agent off, it reads Switched off by Modulify., and its switch, Change with chat and Delete are disabled. No new agents can be added to that site until Modulify switches it back on.

Run history

Each run is kept for 30 days. A skipped run says why: the site is locked, the workspace has used its credits, the daily limit was reached, or the previous run is still going.

Delete an agent

Delete asks Delete "<name>"? first and explains: It stops running and its run history is deleted. Runs it made today still count toward the daily limit. It works on any day, including one it already ran, and it is refused while a run of it is going or its last charges are still settling.

Moving a site

A site cannot be moved to another workspace, transferred or claimed while one of its agents is running or still settling a charge. The move is refused with a message, and you can try again a few minutes later. When a site moves, its agents move with it, and an agent whose owner cannot reach the new workspace switches off.

Over MCP

Creating, changing, listing, switching, running and deleting agents, and reading their runs, are all MCP tools, and the editor chat uses the same ones. See MCP tools.

Next

• Scheduled jobs call a URL on a schedule instead of running an agent.
• Webhooks send your endpoint a request when the site publishes.
• Credits explains what an agent run spends.