Not a chatbot with a to-do list. Five agents, each with a name, a role and a goal, hired in waves and run from one group chat like a small team. Here is what they built in an afternoon, how you steer them, and the list of what was honestly still missing at the end.
There is a version of “AI agents” that is a chatbot with a task list, and there is a version that is closer to staffing. This is about the second one, and it is easiest to explain by doing it: one afternoon, a made-up business, five hires, and one group chat to run them from.
The business was a dropshipping shoe store — everyday sneakers and casual shoes, Australia first, then New Zealand, from a A$25,000 start. Nothing about it is special. That is why it is a fair test.
The workspace has a switch at the top of the left panel: Chat | Agents. Chat is what you expect — ask, get an answer, it waits.
Agents is a different verb. New agent opens a chooser: create an agent, or create a group chat. Creating an agent asks for three things — a title, a role, and a goal — and the moment you save it, it starts working in the background. It appears in the Agents list with a status: Running, Needs you, Finished today. When it finishes, the work is in Doc Manager as a real file, not in a chat bubble.
That is the whole mental shift. You are not talking to something. You are hiring something, giving it a job, and checking its work.
| Hire | Title | Role | Wave | Delivered |
|---|---|---|---|---|
| 1 | Mia · Head of Research | Market research | 1 | A cited two-page opportunity brief, as a Word document |
| 2 | Theo · Brand | Brand designer | 2 | Five names checked for clashes, three logo concepts, a one-page brand board |
| 3 | Priya · Finance | Finance lead | 2 | A business plan and a three-year model with an assumptions sheet and a chart |
| 4 | Sam · Engineering | Web engineer | 3 | A working storefront with a live preview URL and a pre-order form |
| 5 | Lena · Marketing | Marketing lead | 3 | A 30-day content calendar and a drafted launch email |
| 6 | Monday scan | Routine | 4 | A competitor price check, every Monday at 8am, forever |
The waves matter more than the names. A later hire can open an earlier hire’s files, so a finance lead who has read the research brief writes a better model than one starting cold, and an engineer who has the brand board builds a store that looks like the brand. Two hires in the same wave run at the same time — two rows under Running at once, which is the moment the thing stops looking like a chat window.
Mia’s goal, roughly: research this market with sources — the size and growth of online footwear in Australia, the top five competitors and their price points, the realistic dropshipping supplier options with lead times and landed cost per pair, and the three biggest risks. Deliver a two-page opportunity brief as a Word document with every figure cited. Never invent a figure; cite it or flag it.
That last line is the one to copy. It is the difference between a brief you can act on and a brief that reads well. When her row moved from Running to done, the document was in Doc Manager with the citations in it.
Before wave 2, create the group: New agent → Create group chat, call it Launch team, add Mia. Every later hire gets added as they join. The group is an ordinary chat in every way that matters — the same composer, attachments, skills, connectors, your Brand Kit — with two additions. You address members with @Name or @everyone, and beside each member there is a Work button that opens that agent’s own working session.
So the group is where you run the company. The sessions are where the work happens. When you message the group, each member acknowledges first, then does the work, and their replies arrive under their own names the moment they send them.
Theo (brand): read Mia’s brief first. Propose five store names with a one-line rationale each, check each online for an obvious existing business or trade mark clash, recommend one, then generate three logo concepts for the winner and lay out a one-page brand board — lockup, palette with hex values, typeface pairing, three lines of brand voice.
Priya (finance): use the figures in Mia’s brief. Write the business plan as a document. Then build the three-year financial model as a spreadsheet: an assumptions sheet, unit economics per pair (landed cost, shipping, packaging, ad cost, payment fees, margin), a monthly P&L, cash runway from A$25,000, and the break-even month. Add a chart of monthly revenue against cash balance.
Both rows run together. When they finish, the thing to check in Priya’s model is the assumptions sheet: every number downstream should reference it, so changing one assumption moves the whole model. That is what makes it a model rather than a table.
Sam (engineering): use Theo’s brand board and the launch range from Priya’s plan. Build the storefront — home page, a catalogue of the twelve launch products with generated images that match the brand, a product page, and a pre-order form that captures an email. Run your health checks, fix what they find, publish a live preview and give me the URL. Commit as you go.
Lena (marketing): in the brand voice from Theo’s board, create a 30-day launch calendar for Instagram and TikTok — one post per day with a hook, a caption and an image or video prompt. Then draft the launch email to the pre-order list. Draft only; send nothing.
Sam came back with a code project, commits, and a preview URL you can open in another tab, click through, and submit the form on. Lena came back with day one to day thirty and a drafted email that had not gone anywhere, because nothing was connected that could send it. That was deliberate, and it is a habit worth keeping: connect the thing that sends only after you have read what it would send.
In the group: @Sam Change the pre-order button to “Reserve my pair” and republish. His row goes back to Running; the preview updates. In the group again: @Priya Turn the plan and the model into a ten-slide investor deck. A presentation appears, and the numbers in it match her model, because they came from it.
This is what steering means in practice. You do not re-explain the company. You tell an employee to change a thing, and they change it, and the change lands in the file.
New agent, with a schedule — or the New routine button in any agent’s header. Monday scan, every Monday at 8am: check the top three competitors’ prices and new releases and report what changed, with links.
It sits in the Agents list under Routines and listeners. Routines run on a clock; listeners run when something happens. Together they are the part of the company that continues after you stop typing, and they are the reason this is closer to an operation than to a session.
Last message to the group: @everyone Grade what you built against the acceptance checks you wrote before you started, and give me one plain list of what is still missing before this store can take its first paid order.
Each agent had written its checks before starting work — that is how goals work in Agents mode, and the mechanism has its own article. So the grading came back as a list of checks with evidence, and then the missing list:
That list is the most credible thing the team produced, and it is why the wrap should never be cut. An afternoon gets you a real, reviewable draft of a company: the research, the brand, the plan, the model, the store, the calendar, the deck, and a routine already running. It does not get you a trading business, and the agents are the ones who said so.
If a row sits on Running longer than you would like, open its session and say: finish with what you have and list what is missing. If a row says Needs you, it has a question; that is a feature. The Academy’s build your first AI worker is eighteen minutes and covers the single-agent version of all of this; the workers reference is the short form.
Everything above is Agents mode in the Oppermind workspace as it ships today: agents with a title, role and goal; group chats with named members and a Work button per member; routines and listeners; goals graded against checks written up front; Needs you pauses for approval; and deliverables that land as real documents, spreadsheets, decks, designs, images and code projects with a live preview. The Marketplace carries ready-made agent templates if you would rather start from a job description than a blank goal.
Free with no card, then Starter A$9.95, Pro A$29.99 and Pro Plus A$59.99 a month. Media generation is capped per plan, so a team that generates a lot of product images will find the higher tiers matter; unlimited documents, spreadsheets, decks and designs on every plan.
Whatever you’re here to make, make more of it.
Five agents in an afternoon is a demonstration of breadth, not a substitute for judgement. Every figure in Mia’s brief is a figure you check before you rely on it. Priya’s model is only as good as its assumptions sheet, which you should argue with. Theo’s trade mark check is an obvious-clash check, not a search a lawyer would sign. The company they built is a draft, and the reason to build it this way is that a draft you can open, click and correct in an afternoon is worth more than a plan you never finish writing.
Chat is a conversation: you ask, it answers, it waits. Agents mode is staffing: you create an agent with a name, a role and a goal, and it starts working in the background the moment you save it, leaving real files behind. You can create several, put them in a group chat, and run them the way you would run a small team.
It is the office. Every member sits in it, you address them with @Name or @everyone, each one acknowledges and then does the work, and their replies arrive under their own names as they finish. Beside each member is a Work button that opens their working session, so the group is where you run the company and the sessions are where the work gets done.
Because later hires read earlier hires’ files. A finance lead who can open the research brief writes a better model than one starting from a blank page; a web engineer who can open the brand board builds a store that looks like the brand. Hiring in waves is how you give each employee the context the previous one produced — and two agents in the same wave run at the same time.
A cited two-page opportunity brief as a Word document, five store names checked for obvious clashes with three logo concepts and a one-page brand board, a business plan and a three-year financial model as a spreadsheet with an assumptions sheet and a chart, a working storefront with a live preview URL and a pre-order form, a 30-day content calendar, a drafted launch email, a ten-slide investor deck, and a routine that checks competitors every Monday.
No, and the agents said so themselves when asked to grade their work: payments, supplier agreements, an ABN and GST registration, a returns policy, an ads account and a domain were all still missing. There is no checkout tool in an agent’s kit, so the storefront takes pre-orders. An afternoon gets you a real, reviewable draft of a company. It does not get you a trading one.
The business, the figures and the deliverables described are from a demonstration run on a fictional company for illustration; nothing here is business, financial, legal or tax advice, and a real launch requires the items the agents themselves listed as missing, among others. Research, trade mark checks and financial models produced by agents must be verified before reliance. Preview URLs are hosted while a project is active and are not permanent addresses. Names of agents are labels chosen by the user. Oppermind platform prices and plan limits are current as at 13 September 2026, are in Australian dollars, and are subject to the plan terms at checkout.
Agents with a name, a role and a goal, run from one chat, with every deliverable landing as a real file.