Agents – your autonomous team

Agents are AI assistants. Where Chat waits for your brief and Journeys run the automation you designed, an agent owns a piece of your marketing and decides its own next move – it proposes the work, does it, and brings you the results to approve. You stay in charge of what goes live.

Agents are in beta. How many you can have on the team depends on your plan – the lowest plan that includes them runs one specialist, and the plans above it run a team; the hire screen tells you how many you have left. The workspace lives under Agents in the sidebar, with four tabs: Today (the work waiting for you), Team (your roster), Calendar (what the team did and what is scheduled next), and Office (the rules they work under).

Two ways to hire

There are two kinds of agent, and the difference is the same as in a real team:

• A specialist is an employee. It owns one piece of your marketing permanently – your ads, your search visibility, your social – and keeps proposing and doing work in that area for as long as it's on the team.
• A one-off goal is a contractor. You describe a single outcome you want, the agent researches your business, plans the checkpoints, and works until it's done – then it's finished. No standing role, no ongoing cadence.

Both produce drafts into the same Today review queue, obey the same trust levels and spending limits, and show up side by side under your active goals. Hire an employee when an area should never go quiet; hire a contractor when you just need a specific job done.

Hiring a specialist

The Team tab offers eight specialists, each built to own one area:

Every card states what the specialist does, what it needs (connected accounts, some history to learn from), and what it costs to think. You can rename and rebrief any of them, or build a fully custom specialist from scratch. A brand has one of each role – a role already on its team is not offered again.

After hiring, the specialist reviews your setup and asks whatever it still needs – one question per screen, then Save and let them start. If that first read stops part-way, its profile says Setup did not finish with Try again.

You don't have to wait for it. Your bell tells you how setup ended: the question it needs answered before it starts, that setup did not finish, or that it is on the team (that notice opens its first read of the brand). Agents › Today counts the questions it still needs answered before it starts; optional ones are not counted. When a batch of its ideas is waiting for someone to pick one, the bell says so too. A batch that started a goal by itself is not announced as ideas: the goal's own notices (a plan ready, a question) go to whoever runs Agents for the brand.

Who is told:

• How setup ended: whoever runs Agents for the brand, the person who hired the teammate, and whoever pressed Try again.
• Ideas waiting for a pick: whoever runs Agents for the brand, and whoever pressed Bring me fresh ideas. The person who hired the teammate is told about its first batch.

Anyone other than whoever runs Agents is told only if they can still open the brand when the notice is sent. A notice shows while its brand is the one open in the app. When a teammate is let go or deleted, its notices go with it.

Hiring from the chat

You can also ask the chat for a teammate – "hire us an SEO agent", "do our ads every week", in any language. The reply names the teammate that fits and draws a card under it: what the teammate does and needs, how freely Agents run on the brand (manual review, countdown or auto-publish), what setting up and its ideas cost, and who is told how setup ends. Hire puts it on the team – the same hire as on the Team tab, recorded in Activity as hired "from the chat". The card then follows its setup: on the team, questions to answer on its profile, or Try again if setup stopped.

The chat offers the area specialists and the Companion. The Autonomous Customer Support Employee and custom specialists are hired here on Agents. When Agents are not switched on for the brand, the card says so and opens Agents for it instead; when they are paused, you can hire and the teammate starts once they are switched back on.

In an organization, only the people an administrator has ticked May hire AI teammates for a brand (Team, in the admin portal) can hire on it – on Agents and from the chat alike; the account of record always can. When you are not ticked for the brand, the chat says so and names who can tick you, with no card; on Agents, Hire says the same.

A teammate who was let go comes back only through the same checks as a new hire: the tick, one per role and a free place on your plan.

The Autonomous Customer Support Employee

One teammate on the same roster is not a planner: the Autonomous Customer Support Employee proposes no work of its own and owns no goals. Its work arrives from outside – every message a customer sends your helpdesk – and it is hired for one brand at a time, on the brand whose desk it should work.

It also earns its job in three steps, and every move between them is yours to make:

1. Learning – works silently and learns your desk. Nothing reaches a customer and nothing appears in your conversations.
2. Copilot – proposes every reply, a person sends it.
3. On its own, case by case – answers the kinds of question an administrator has switched on, one at a time, and everything else stays a proposal waiting for a person.

Its profile shows which step it is on, the numbers behind what it has compared, and the sentence in which it says whether it believes it is ready for the next one. It never moves itself and it never switches a kind of question on: you decide, on a button there, and you can move it back at any time. Before any of it runs, an administrator of your organisation confirms the AI Support Agent Addendum in Settings › Agents, where every agreement your agents work under is listed first.

The desk itself – the modes on a conversation, what the employee may answer from, and the questions it could not answer – is in the Helpdesk manual.

Training

Its profile has a Training tab, and no other teammate does – the others are taught on their Knowledge tab (see Teaching a teammate). Some of what a desk knows is not written down anywhere – it is a path through a screen somebody learned by watching a colleague. Training is how you show it that one: record your screen while you do the task, talking as you go, or add a few screenshots. It writes down the steps it understood, asks about whatever it was unsure of, and you read them back beside the pictures it read them from, fix anything it got wrong, and approve a version.

Two things to know before you start. Use a test customer, because a picture cannot have names and numbers taken out of it the way text can. And what it learns is kept and versioned but is not used in a reply or a lookup yet – this step exists so you can see what it makes of your own recordings first.

What to record, what it costs, what happens to your recording when you approve, and how versions work are all in the Helpdesk manual, under Training: showing it how something is done.

Handing a project task to an AI teammate

Any task on a project board can be assigned to a member of your AI team the way it is assigned to a person: open the task, choose the teammate under Assignee. The Companion is the teammate built for this – it has no agenda of its own and only does what it is handed – but any ready specialist on the brand can take a task in its area.

The teammate plans the fewest steps that deliver exactly that task, does the work, and brings the result back: the task's record shows what was delivered, the first piece becomes the task's linked content, and the task moves to Done when the work is done. While it works, the task panel says what is happening – planning, working, waiting for your approval – with a door to the goal behind it. Take the task back at any time by changing the assignee; the work stops.

A handed-over task is one goal, billed like any goal: planning, then the tools it uses. The Office rails apply – a paused brand or a spent budget refuses the hand-off and says why.

Hiring a one-off goal

For a single job, choose hire a one-off goal on the Team tab and describe what you want done in plain words. The agent researches your business, plans the checkpoints, and starts working – you can edit or pause it at any time. Its drafts arrive in Today like everyone else's, its goal sits in the active-goals list alongside your specialists' work, and when the outcome is reached, the job is done.

While a plan is being built

Planning a goal takes a couple of minutes, whether you started it yourself or promoted one of a specialist's ideas. The agent reads your brand, maps the checkpoints, and then puts the plan in front of three reviewers who check its scope, its voice, and whether every step can actually be measured.

The goal appears in your list straight away and shows what the agent is working on right now. You do not have to wait on the page. Close the tab, switch brands, or come back tomorrow – the work carries on without you, and you get a notification in the bell when the plan is ready. Open the goal at any point to see how far it has got.

Two things can happen instead of a finished plan:

• The reviewers hold it. If the plan scores badly on something that matters, it waits for you rather than running unattended. The goal says why, and every checkpoint is saved – press Approve and start if you are happy with it anyway. You can turn this off per brand in the Office if you would rather plans always start on their own. When the agent left questions open, they sit under Questions the agent could ask in Plan details, each with an answer box: answer the ones you can and press Answer and plan again. Your answers are saved for the whole team, and a new plan is written with them in place of this one – the page opens the new plan while it is being written, and this one goes only once the new one exists. Approving brings the time the plan starts: Approved. The plan starts tomorrow at 09:00.
• Planning stops. Occasionally a plan cannot be generated. The goal says so in plain words, you get a notification, and nothing is lost – the idea stays in the specialist's list to try again.

How they work

On their own cadence – roughly weekly, or on demand with Bring me fresh ideas – each specialist proposes a handful of goals from your data, each viewed through a different lens (compounding wins, new capabilities, research, creation, optimization). What happens next depends on the trust you've given them:

• Asks me before publishing – ideas and drafts queue for you; nothing ships until you approve.
• Waits 24h, then publishes – drafts auto-publish a day after creation unless you reject them first.
• Publishes immediately – work goes live as it's produced; you're notified after.

Repeating goals follow the same setting. Each run's posts become drafts that wait for you, or publish a day later, unless the team publishes immediately. A change in one of your other systems – an ad budget, a CRM record – runs on its own only when the team publishes immediately; otherwise it waits for your approval.

The same rule covers every step a goal takes, not only posts. A step that would publish somewhere, change an ad, reply to someone, fill in a form or change something on a website stops and waits in Needs you – with the exact action and what Approve will do – unless the team publishes immediately. A browser step that only reads (research, monitoring) runs on its own. A browser step that would act shows the task the browser is given, word for word (the first lines, the rest under More), the page it starts at, the page it acts on, and the site: approving lets a browser do it there, through the Chrome extension of someone on your team with the app open, signed in as them, or else the cloud browser may try once with a saved login. A setting for one platform, such as asking first before posting on X, holds even when the team otherwise publishes immediately. When you approve a piece of work that a later step will publish, the card says so.

When a goal makes a video. A video with words is made the way Video Studio makes an ad from a brief: a script with a line for every scene, one voice for the whole video, captions and music. Who says the words follows the step: a voice-over with nobody on screen when the step asks for no people (or names a voiceover or a narrator), your brand's own actor when it asks for a presenter, and for a brand that sells to companies a voice-over unless a presenter is asked for. A short clip with no words (a product spin, a background loop) is made as a plain clip. A video takes several minutes to render after its step is done, so the step that publishes it waits for the file, without counting a failed attempt, and the team's other goals carry on meanwhile. If the render does not finish, that step stops and asks you, with the reason and a link to the video in Video Studio: nothing is posted without it. YouTube, Instagram (as a Reel), Snapchat and Reddit steps post the video. A LinkedIn, Facebook or X step posts the words, and its receipt says the video did not go with them.

Trust is set globally in the Office and can be overridden per specialist. Everything they produce lands in the Today tab. When a goal is finished – or has finished setting up something that now runs on a schedule, like a Monday report – it hands the work over at the very top: Here is your …, who did it, when the next one comes, and the thing itself, with Open it and Got it. The bell and the teammate's e-mail say the same and open it for you. Below that, Today is two lists:

• Needs you – everything the team is waiting on a person for, with the button that resolves it on the card: a question a step stopped on (Retry, Skip this step, or an answer in your own words – and, when that is what stopped it, Add credits, Raise the daily limit or Add the login), a plan the reviewers held (Approve and start), a login it needs (Sign in now), a piece of work waiting for your approval (Approve or Reject, with a note), a specialist still setting up, a notice to dismiss. A limit that has stopped the team comes before everything else, then questions. The headline at the top of the page reads from the same list, so it never says the team is choosing what to do while something waits on you.
• Done for you – everything the team has made, newest first, shown as the thing itself: the title, a miniature or an excerpt, who made it, for which goal, at which step. Open one and it slides over the page – the same reader as the Library, with Discuss, copy and download – and a link takes you to the goal it came from. The Library keeps all of it; Today is the desk.

Below the two lists, Recent activity is the diary: what ran and when, with the agent's reasoning behind "Why".

When things happen

Steps run in order, each at its planned time or as soon as the step before it is done, whichever is later – and a piece waiting for your approval holds the steps after it. The team also sleeps through your quiet hours and the days it does not work (both set in the Office), so a step planned for then runs when the team is back.

The pages say so in your own clock:

• On the goal, each step that has not run reads runs Tue 15:30 (the earliest it can run), next (it runs now), then (straight after the one before it) or waiting for you (a piece of it is in Needs you). The line under the steps says the same: Earliest times. Steps run in order, so a piece waiting for your yes holds the ones after it.
• In progress on Today, a goal waiting on its next time reads Next: Write the post, tomorrow at 15:30, one waiting on you reads Waiting for you, and Working means a step is due now.
• On a piece of work, the card says when the later step that publishes it runs – no earlier than that time, and once the steps before it are done when there are any.
• The Calendar places upcoming steps at the same times.

The Office: rails they can't cross

• Master switch – pause the whole autonomous operation for a brand with one toggle.
• Spending limits – credits per day, credits per week, and drafts per day (50 for a new brand). When a limit stops the team while work is waiting, you are told once – on your bell and at the top of Needs you – with which goals wait, when the team picks up again (after the limit resets and outside quiet hours), when each goal finishes at this limit, and whether a deadline or a goal's first scheduled run moves because of it. Raise the limit … and continue sets the limit to enough for what is waiting and the team goes on within minutes; Set the limits opens this card; Leave it until it resets keeps things as they are. A scheduled run a limit holds – a Monday brief, say – runs when the limit resets, or within minutes of a raise, rather than skipping to the next week. These limits never touch your own manual usage.
• Working hours – active days, quiet hours, and timezone. Nothing publishes at 3am unless you want it to.
• Credential vault – saved sign-ins (encrypted, 2FA supported) for the platforms your agents work in, used only when an agent needs to log in for you.
• Trigger rules – "when a condition is met, your team does something automatically."
• Ledger – every action the team took, who took it, and the trust level and budget it ran under, newest first. Changes to the rules above land here too, with who changed them. Export it as CSV when you need the evidence outside the product.

A note on Bring me fresh ideas: pressing it yourself skips the working hours, but not the budgets. A batch costs credits, and the daily and weekly limits are your ceiling on what the team spends whether the team or you started it.

The SEO Strategist reads your search position first

Before it proposes anything, the SEO Strategist reads where your site stands in Google today, in the market you sell in:

• every search your site ranks for in the top 100, the page that ranks, and how it moved since the last check
• the searches you told it you want to own, measured – including the ones nobody searches for, which it will say plainly
• the searches your buyers make for your category, and the comparison searches around the competitors you named ("x alternative", "x vs y", "x pricing")
• the sites that compete with you in Google, and what the closest of them win where you do not rank
• the live results page for the searches that matter most: who wins, the kind of page that wins, whether Google shows an AI Overview and whom it cites, and the questions people also ask
• a technical check of your most important pages, and your sitemap, so it never proposes a page you already have
• your site's latest site audit: problems by kind, broken links and page speed on the key pages. When there is none, or it is more than 14 days old, it starts a new crawl and reads it on its next round

Each idea names the pages and searches it works on, and the figures it rests on appear under it – search volume, difficulty, your position. Those figures come from Google search data and your own site, never from the AI. When you make an idea a goal, the plan starts from the same research.

No connection is needed for any of this. Connect Google Search Console in Integrations and it also reads what your site actually earns in Google: clicks, impressions and average position for each search and page over the last 28 days against the 28 before, your sitemaps, and how Google indexes your key pages – including when Google picked a different canonical than the one your page declares. It only reads; it never changes anything in Search Console.

A goal the SEO Strategist works on shows its pages in Google search: each page's clicks, impressions and average position in the 28 days before the goal started against the latest 28 days. These are Google's numbers for the page – they measure the page, not the cause – and until 28 days have passed the page says when the first clean comparison is due.

When the brand has an app saved on the App Store Optimization page or in Brand settings, the SEO Strategist also works on its App Store and Google Play listings: its plans can include a new listing from App Store Optimization, read from the stores the same way the page reads them. It writes the listing and saves it to your library; it never changes a listing in App Store Connect or the Google Play Console and never replies to a store review – a person sends those. A brand with no app is never asked for one.

The Social Media Manager learns your channels

The social specialist keeps a playbook per platform – what wins on LinkedIn often flops on X, so rules never cross platforms. You rate each published post (Worked / Meh / Flop); once a platform has three rated posts, the weekly reflection proposes rules – what works, what doesn't, voice, timing, formats – and you accept or reject each one. Accepted rules shape everything it makes next. You also choose its mode: original posts, joining conversations, or both.

Teaching a teammate

Every teammate except the support employee can be taught from your own material, on the Knowledge tab of its profile, under What you taught it. It is optional – a teammate works without it.

What it can take from your material. You choose one of four:

• How it should sound – your tone, the words you use and avoid, and pieces that sound right.
• How the work is laid out – the sections of a report or an e-mail in order, or a spreadsheet's columns.
• What to always or never do – rules it keeps every time.
• What it should know – facts about your products, prices and customers, or the steps of a task in one of your tools.

The material. Add a file (a PDF, Word, PowerPoint, Excel, CSV or text file, up to 10 MB), choose up to five pieces from your Library, add a web page, or write it in the box. A spreadsheet given as a layout is read for its column names and its first two rows only; given as anything else, every row is read, so leave customers' details out. E-mail addresses, phone, card and account numbers are taken out automatically before it reads anything.

It studies, you check. It writes down what it took from the material: what it will do from now on in its own words, the rules it will follow, the pieces it keeps word for word (for their voice and shape – it never copies their facts, names, numbers or offers), the layout or columns, what it now knows, and what it was not sure of. Nothing counts until you press Use this. You can edit any part first, and edit it later – saving is your approval. Put it away stops it being used; Bring it back restores it.

Where it is used. From its next piece of work, in everything it writes for the work you chose – the plan, each step, the tool that writes the piece, and scheduled runs. A layout replaces the tool's own structure, so a report taught as a spreadsheet comes back as a table with exactly those columns, which you can download with Export CSV. The Social Media Manager can keep a lesson to one platform. Each lesson shows how many pieces of work it has shaped, and a result made with one says so on its Made by line. The brand's guardrails always come first. Teach the same to copies a lesson you are using to another teammate on the same brand, at no charge.

What it costs. Studying is priced on what it reads, and the form shows the most it can cost before you press Teach it – the rest is given back once it has read, and all of it if it cannot be studied. Checking, editing, putting away and copying are free.

When an agent needs an answer

A plan can include a step whose deliverable is a question for you – which product to feature, which offer to run – and a step that fails three times, or costs more than the day allows, stops and asks too. The goal waits there, paused or not. The question rings your bell, sits at the top of Needs you on Today and on the goal page with an answer box, and a specialist also emails it when you have email delivery on; a reply to that email counts the same, and so does a message to the agent in the app (see below). It is written in plain words: the step by name, what went wrong, and the three answers it understands – "retry", "skip", or what the step needs. Your answer becomes part of the agent's instructions and the goal continues. Nothing is published by an answer – approval stays where it was.

Agents do not ask you what your data already answers. Every step sees what your warehouse holds, and a question such as "where does the MRR data live?" is checked against it before it reaches you – if your data answers it, the agent carries on. Where a step needs a definition ("trials" when you record signups), the agent uses the one your data supports and says so in its summary; where something is not in your data, the step delivers the rest and names the gap. The questions that still reach you are the ones only you can settle – a target, a choice between real options, a login, an approval – and a step that keeps failing.

The goal page also lists What it needs from you – every login the plan named and the brand does not have yet, with the button to add it, and every connection – and What it will deliver: the mix of work the plan produces, which pieces are done, how many steps go live.

Talking to an agent

Every specialist has a Talk to them button on its page, and every question card and goal page has Ask. It opens a conversation with that teammate, in its own voice, about its work – where a goal stands, why it chose a step, what it needs from you. Write "status" for where things stand, "pause" or "resume" to stop or restart it. A message about a goal that is waiting on your answer is taken as the answer, and the agent says so. Everyone on the brand sees the conversation; each message carries who wrote it.

An answer the agent writes costs 1 credit; status, pause and resume are not charged. It cannot approve a plan, approve a draft or publish anything from the conversation – those stay on the goal page and under Needs you. The e-mail address under its name still works the same way, and both land in the same record.

Replying and engaging happens through your saved logins or the built-in cloud browser session – so it works on platforms without a posting API, and every session is visible in the activity log.

Watching them work

Each specialist's profile carries an Activity timeline – goals started and completed, drafts published (with links to the live posts where the platform provides them), your verdicts, and playbook rules learned – and a Conversations tab with what you said to it in the app, what it e-mailed you and what it heard in meetings. The roster shows at a glance who's working, who's paused, and who's waiting on an answer from you. Pause, resume, or let a specialist go at any time. On a goal page, every finished step shows the result it produced by its own title, with Open to read it there.

What it costs

Planning a goal is priced on what it uses – usually around 30 credits, charged once the plan is written; a plan that fails costs nothing. A batch of fresh ideas costs 2 credits. The SEO Strategist's batch is priced on what it uses – usually about 5 credits, and about 10 when it refreshes its read of your search data, which it does at most once a week. A batch that fails costs nothing. Executing goals bills per tool used – each tool's credit cost is shown on its own page in the product – plus what the agent's own thinking uses as it works, at the same rate as chat. A cloud-browser session, used for platforms without a native API, reserves up to 40 credits and settles on what it used. A video is priced like the same video in Video Studio: a 20-second scripted Short on the default model is about 58 credits (the script, three scenes, captions, music and the voice), and the plan shows that figure before the step runs. Your daily and weekly limits in the Office are the hard ceiling; agents stop before they pass them.

Agents, Journeys, or Chat?

• Chat – you drive, one brief at a time.
• Journeys – you design the automation; it runs exactly that, forever.
• Agents – you set the goal and the rails; they decide the moves.

They compose: an agent can build journeys, and everything any of them produce respects the same suppression, unsubscribe, and frequency rules as work you ship yourself.