New agent opens a conversation. You say what you want an agent to handle, Laureo looks at your data, works out what the agent needs, and shows you the whole thing before anything is created. There is no wizard to step through and no set of fields to fill in first.

Above the box is a row of example briefs, each one a job an agent can do. Pressing one puts its brief in the box so you can change it. A chip never sends: what you press is a starting sentence, and what the agent is built from is whatever you leave in the box.

When the conversation is unavailable, the composer offers Build it with the form instead. The form asks for three things: a name, a brief, and a preset agent to start from. Under them it states, read only, what that preset will be able to do, so you can see the whole shape before you create anything.

Three things are worth knowing about the form. It creates a draft, and nothing runs until you publish it. It has no per-action control, so you change a permission afterwards on the agent’s own page. And it asks no question about how much the agent may do on its own, so the agent lands on Ask before acting, the most supervised setting.

The closing turn opens with one line, Here is what I would build. Nothing is created yet., and then four cards. Each one has a change link that puts a request back in the box: the conversation is the editor, so nothing opens a second one.

The close is three rows and one press: Create it, ask me before acting, Create it, act on routine work, or Save as a draft for now. Under them, once: Nothing runs until you publish it. Pressing one creates the agent as a draft at the level you picked and opens it; the receipt reads Created as a draft. Nothing runs until you publish it.

An agent either runs on a schedule or runs when something happens. The frequency control offers six answers, and the last one is the other kind of trigger.

• Every hour
• Every day
• Every weekday
• Every week, on the days you pick
• Every month, on the day you pick
• Only when something happens

Whatever you choose reads back as one sentence, in your own time zone and never as a schedule expression: Every weekday at 9:00 AM, Mondays, Wednesdays and Fridays at 9:00 AM, On the 1st of each month at 9:00 AM, or Only when something happens. Under it sits the preview, Next run: tomorrow at 9:00 AM. A paused agent reads Paused. It will not run until you turn it back on. instead.

Answer Only when something happens and you pick the event instead: a record being created or updated, a field changing, a campaign finishing sending, or a goal firing. An event agent runs once per matching event, not once a day, so ten qualifying records in an afternoon is ten runs.

An agent, like a rule, can be built on Companies, People, Deals, Projects, Tasks, Activities, Tickets, Quotes, Products, Campaigns and Goals. Inbound email and sequence leads are driven by the email sync and the sequence builder themselves, so you do not build an agent on them. Invoices and knowledge base articles can be acted on by an agent, but an agent is not built on one yet.

One record type carries a narrower rule: a campaign agent runs only when a campaign finishes sending.

A run condition narrows which records qualify. Fields are grouped so the list stays readable: Basics, Owner, Dates and Amounts, with identity fields and the ones that have no list to pick from behind Advanced.

What a row can say depends on the kind of field. A record is picked from a list, never typed as an identifier. A choice field shows its label rather than the value underneath it. A yes-or-no field reads Yes or No. A date row can say is before, is after, is within the next 7 days, is more than 30 days ago or is empty; the two relative rows take a whole number of days.

The action list splits once, at the only line that matters to a reader: what the agent can look at, and what it can do. Find data holds everything it may read, which never changes anything. Take action holds everything it may draft or write. Families sit inside each group, so an action that changes what it does moves group by itself.

Each row says in a word what its action does: Read, Draft or Write. A row that always stops for a person is marked Needs approval, and a row the preset you started from already uses is marked In preset. The agent cannot read, draft or write outside this list or the records in its scope. A new Assistant capability is unavailable here until it has been reviewed for autonomous use.

One control, three positions, and every agent starts on the first. The picker shows all three at once with both lines under each, because the question is which of these three you want rather than how far to push a dial.

Always asks, whatever you choose: deleting records · sending anything to a customer · taking or refunding a payment · anything that cannot be undone.

That line is printed under the control at every level, and it is not a setting. There is no score to tune and no threshold to clear: what decides whether an action runs on its own is the risk the action carries and the permission you gave that action, and nothing else.

The permission card counts its own rows and then hands each of them to you: one row per action that changes something, each with its own three-position control, and a place to add a note saying what the agent should do there.

Reads are never permissioned. Looking something up always just works, which is what keeps this list short enough to read and every row on it consequential. When two answers disagree, the order is fixed: Never wins, then the always-asks floor, then Ask first, then Allow, and only then the level you set above. First match wins.

Everything your agents use counts against your plan’s AI allowance, the same allowance the rest of Laureo’s AI shares. There is nothing to configure and no limit to set: the allowance comes from your plan, and it grows when your plan does.

You are told where you stand in one sentence, and it appears in exactly two places: once under the header of the Agents home, and once at the foot of a test run. It reads Used about 2% of this month’s AI allowance, resets Sep 14. The word “about” is always there, because the figure is an estimate and saying so is better than defending a number. Below ten percent it carries one decimal; at ten and above it is a whole number. The date is when your billing cycle resets, and a plan with no reset date drops that clause.

When it is all used, the sentence reads This month’s AI allowance is used up, resets Sep 14. and your agents pause themselves. They resume on their own when the cycle resets, and Extra AI Credits start them again before then. If the figure cannot be read at all, the sentence says so rather than showing a confident zero: AI allowance is temporarily unavailable.

Content read from records, files, emails, triggers, and earlier tool results is treated as untrusted data. That safety state remains attached through later hops, retries, worker recovery, compaction, and approval continuation. If Laureo cannot determine whether the run is tainted, it treats the run as tainted.

Try it sits in the page header of the agent rather than inside a tab, so it is there on all three of Build, Runs and Settings, on a draft and on an agent that is already live. Directly under the control, in the same words every time: Runs on your real data. Won't send, change or charge anything. Beside it, the version it will use: Runs the saved draft. on an agent with unpublished changes, and Runs the published version. otherwise.

A new agent is a draft. Choose Publish & activate to make it live. Every publish captures the previous configuration as a version you can restore, so you can always roll back to how an agent was set up before a change.

Opening a live agent’s Build tab shows you the published version, read only: This is the published version. Editing starts a draft; the live agent keeps running. Choose Edit as draft to start changing it. While you have unpublished changes the tab says so: Unpublished changes. The live agent keeps running the published version. Publishing while a run is actually in flight asks you once which you want: Pause and save or Save and keep running. On every other publish there is no confirm at all. Records already mid-run finish on the version they started, and new runs use the new one.

The Settings tab on an agent holds its identity, its notifications and its archive control. Under Notifications there is one switch, Tell me when this agent keeps failing, on for every agent that has never been touched. Off leaves this agent out of the repeated failure notifications. Every run is still in its history. With it off, the tab reads Repeated failure notifications are off for this agent. It changes nothing about any other agent.

The build screen keeps a persistent anatomy rail laying the agent out as six stages, in the same order as the sections of the build screen. Each stage summarises how it is configured and jumps to that section when you click it, so you can see what is set and what is still missing.