What is an AI office? How your AI agents work as one team

An AI office is one shared place where a business's people and AI agents work together as a single team. Each person has a chief, an agent that runs their other agents. Teammates are agents hired for whole jobs, and each one can run on the engine that suits it, such as Claude Code, Codex or Vestra's own. Everyone works from one shared memory, under one set of rules for money and approvals, on an office floor you can see. It exists because today's AI agents, from Claude Code and Codex to Grok Bot, Meta Muse, OpenAI's dot and Hermes, are built for one person and work alone.

Quick answer

• What it is: a workplace for people and AI agents, where agents are hired for whole jobs and work as one team.
• Why it matters: personal AI agents can't see each other, so people end up pasting context between them, carrying work from one to the next and chasing every result.
• What it contains: a chief for each person, teammates on the engines you choose, a kit for every agent, a shared brain, a visible office floor, spaces, shared apps, access control and rules that always ask before money, deletions or first contact.
• Who it's for: any business that wants AI to do real work across sales, finance, support, marketing and engineering, not just help one person at a time.

Key takeaways

• Most AI agents today are personal: one person, one account, one app, one company's AI.
• Personal agents are good at their jobs. The problem is that they work alone, so people become the glue between them.
• An AI office turns those agents into teammates: hired for whole jobs, managed by a chief, sharing one memory and one rulebook.
• In an AI office you can mix engines. One teammate writes code on Claude Code, another reviews it on Codex, and a third answers customers on Vestra's own engine.
• A few things should always wait for a person: spending over a limit, deleting something and contacting someone new.
• Start with one job that eats your week, check the receipts for two weeks, then hire the next teammate.

Contents

• What is an AI office?
• Why do AI agents need an office?
• What's the difference between a personal AI agent and an AI office?
• What are the benefits of an AI office?
• Who is an AI office for?
• How is an AI office different from automation tools like Zapier?
• Chatbot, AI agent, AI employee or AI office: which is which?
• What are the parts of an AI office?
• How does one piece of work move through an AI office?
• What does a week in an AI office look like?
• What does an AI office look like for different businesses?
• Which jobs should you give an AI office first?
• What mistakes do teams make when moving from personal agents to an AI office?
• How do you tell a real AI office from a renamed chatbot?
• How do you set up your first AI office?
• How much does an AI office cost?
• Is an AI office safe?
• What doesn't an AI office do?
• How does Vestra build the AI office?
• FAQ
• Glossary

What is an AI office?

An AI office is a shared workplace, usually shown on screen as a map you can walk around, where every person and every AI agent in a business has a desk. Agents in it are teammates hired for whole jobs. They use their own computers, phones, inboxes and the business tools the team already runs on. They read from one shared memory and work under one set of rules.

The idea is simple: give AI agents what people already have when they work together. People have offices, channels, meetings, managers and a company handbook. Most AI agents have none of that. They live inside one person's chat window and forget the business the moment the conversation ends.

An AI office has four defining traits.

1. It's built for a team, not one person. Several people and many agents work in it at once. Each person has their own chief, and the chiefs talk to each other.
2. Agents hold jobs, not prompts. A teammate is hired for something like "get us paid on time" and owns it for months, with goals you can check.
3. Everyone shares context. There is one memory for the whole team, with a source for every line, so what one agent learns reaches the rest.
4. One set of rules covers everything. Spending caps, approvals, logins and receipts work the same for every agent, whatever engine it runs on.

A good test is to ask who is doing the managing. If you are briefing each agent, checking each result and copying work from one tool to another, you have a set of personal agents. If a chief hands out the work, checks it and brings you only the real decisions, you have an office.

Where the term comes from

"Office" is a deliberate word. A workplace is where a team's context lives: who does what, what was agreed with each customer, what went wrong last time and who can approve a discount. When agents have a place like that, they stop being clever tools and start being colleagues. Vestra's handbook puts it in one line: Vestra is one office for your people and your AI agents.

Why do AI agents need an office?

AI agents need an office because they got good at doing work but still work alone. Each one knows only what its own person told it and can't see what a colleague's agents are doing. So people have quietly become the glue between agents. An office gives people and agents one place to work as one team.

Vestra's handbook names four jobs that people end up doing for their agents. If any of these sound familiar, you have the problem an AI office solves.

Brief: the same context, pasted into every agent by hand

Every personal agent starts from its own blank page. When a customer changes their mind, you tell your writing agent, then your support agent, then your developer's coding agent. Each one needs the same background, the same names and the same rules, typed again.

Carry: work moves only when a person copies it

A research agent finds three rival price changes. A sales agent should use them on today's calls. In a world of personal agents, that only happens if a person copies the findings out of one tool and pastes them into another. Work sits still until someone carries it.

Chase: "Is it done?" and then "Is it right?"

Personal agents report back to the person who asked. Nobody else knows the state of the work. So managers spend their mornings chasing: did the invoice go out, did the client reply, did the fix pass its tests? And when an agent says it's done, someone still has to check that it's right.

Click: forms, settings and dashboards for every request

Every tool has its own settings page, its own approval screen and its own billing dashboard. Ten agents means ten places to look and ten sets of rules to keep in line.

What an office changes

An office replaces each of those four jobs with something better.

• Brief once. Context travels with the work, minus anything private, to every agent that needs it.
• Hand off. Chiefs pass work between people's agents directly, without anyone copying and pasting.
• Checked. Work counts as done only when something other than the agent that did it has checked it: a test, the source, a second agent or a person.
• Just talk. You say what you need, in a sentence, to your chief. The screens are there to see the work, not to drive it.

What's the difference between a personal AI agent and an AI office?

A personal AI agent works for one person, in one app, usually on one company's AI. An AI office brings many agents together for a whole business: a chief runs them, they can run on different engines, they share one memory and one set of rules, and you can see all of them at work in one place.

The personal agents most businesses already use

Here is what most people mean by "AI agent" today. Every one of these is a strong product. Look at who each is built around.

• Claude Code is Anthropic's coding agent. It reads a codebase, edits files, runs commands and opens pull requests, in the terminal, IDEs, the desktop app, the web and Slack. Its core experience is one developer working in one repository.
• Codex is OpenAI's coding agent, used by developers inside the ChatGPT world.
• OpenAI's dot starts you with one primary dot inside ChatGPT that takes on ongoing work. Specialist dots for company roles are in focused enterprise pilots. A dot can't have its own email address at launch, and you can't view or delete its memories one by one.
• Grok Bot lets you create Bots one by one in xAI's app, each briefed in your first message. Your Bots share one cloud computer and one claimed email address, and there's no settings page for a personal Bot.
• Meta Muse is, in Meta's words, a personal AI agent built for everyone and around your whole life. It runs on Meta's own model and is available in the US and Canada.
• Hermes Agent from Nous Research is an open-source agent you host yourself. It builds a deepening model of who you are across sessions and can run named specialist Bots that you set up and look after.
• Perplexity Teammates are agents inside Perplexity Computer that need an individual Perplexity Max subscription at $200 a month, set up one by one in your sidebar.

Each of these starts from one person, one account and, in most cases, one company's models. That's the right shape for a personal assistant. It's the wrong shape for a business that wants all of its AI to work together.

What goes wrong when every agent works alone

When a business adds personal agents one person at a time, the same problems show up within weeks.

• You become the manager of every agent. Each one needs its own brief, check-ins and corrections.
• Nothing is shared. What your developer's agent learned about a customer never reaches your sales agent.
• Work falls between them. No agent can hand a job to another person's agent, so work waits for a human to carry it.
• The rules don't match. Each tool has its own bill, its own spending controls and its own idea of what needs approval.
• Nobody can see the whole picture. There is no single place that shows what every agent is doing right now, what's stuck and on whom.
• Costs stack per person and per tool. Five people each paying for two personal agents is ten subscriptions, ten allowances and ten sets of limits.

Personal agents vs an AI office, side by side

• Built for: a personal agent serves one person; an AI office serves the whole business, people and agents together.
• Who manages it: with personal agents, you manage each one yourself; in an office, your chief runs your agents and brings you only real decisions.
• Which AI does the work: a personal agent runs on its maker's AI; in an office, each teammate runs on the engine you pick, and you can mix engines on one team.
• Memory: a personal agent remembers one person's chats; an office keeps one shared brain, with a source for every line, that every agent reads before it acts.
• Hand-offs: personal agents need a person to carry work between them; in an office, chiefs hand work to each other directly, with the context attached.
• Rules and spending: personal agents each have their own limits; an office has one set of caps, one list of things that always need a yes and one receipt per job.
• Visibility: personal agents live in separate apps; an office shows every desk, every agent and what each is doing, live.
• Colleagues and clients: personal agents rarely talk to other people's agents; in an office, chiefs work together inside the company and with clients' chiefs outside it.

Can you keep using the agents you already have?

Yes, and that's the point. An AI office isn't a replacement for Claude Code or Codex. In Vestra, a teammate can run on Claude Code or Codex as its engine, so your engineers keep the tools they trust, and those agents become teammates: with a desk, a job, the shared memory and the office rules. Vestra's own engine works with the top models, and the engine picker also lists Gemini, Grok and Muse.

What are the benefits of an AI office?

The main benefit of an AI office is that work moves between agents and people without anyone carrying it. The other benefits follow from that.

• Less managing. You talk to one chief instead of briefing every agent, and you hear only about real decisions.
• Context that compounds. Every finished task, customer reply and correction goes into one sourced memory, so the whole team gets better every week instead of each agent separately.
• The right engine for each job. Coding, reviewing, writing and research can each run on the engine that does them best, inside one team.
• Fewer interruptions across time zones. Colleagues' chiefs answer what their people have shared, so a question asked at 9am in Austin doesn't wake anyone in Tokyo.
• Work you can trust as done. Work counts as done only when it has been checked, and the proof travels with it.
• One set of rules. Spending caps, approvals, sealed logins and receipts work the same for every agent.
• One view of the business. The office floor and the Right now feed show what moved, what's stuck and on whom, without a status meeting.
• A single bill. One plan covers the office, instead of a stack of per-person subscriptions.

Who is an AI office for?

An AI office suits any business where work crosses more than one person or more than one tool. Vestra's handbook describes offices for five kinds of user, each starting from the same office.

• Personal use: one person with a small team of agents for life admin and side projects, usually on the Lite plan.
• Solopreneurs and consultants: one person running a business like a much bigger company, with teammates for email, proposals and money.
• Small and mid-sized businesses: a team whose agents take on the busywork of a growing business, such as invoicing, customer care and follow-up.
• Startups: a founding team that wants engineering, sales and support teammates from day one, without the payroll.
• Enterprises: companies that want security, controls and a team for every department, run in Vestra's cloud, their own cloud or on their premises.

If only one person in your business uses AI, and only for their own work, a personal agent may be all you need. Once two or more people rely on agents, or one agent's work needs to reach another, an office starts paying for itself in time not spent briefing, carrying and chasing.

How is an AI office different from automation tools like Zapier?

Automation tools follow fixed rules you build step by step: when this happens, do that. They're excellent when the input and the right response never change. An AI office works from goals. A teammate decides the steps, handles the cases your rules didn't foresee, uses websites that have no connection, and asks when it's unsure.

The two fit together. Keep predictable steps, like copying a form into a spreadsheet, in ordinary automation. Give a teammate the parts that need judgment, like deciding which overdue clients to nudge and how, or reading a customer's email and filing the real problem. In an AI office, a routine can also start the moment something happens in a connected tool, so many simple automations become part of a teammate's job, with the same rules and receipts as everything else.

Chatbot, AI agent, AI employee or AI office: which is which?

These terms overlap and vendors use them loosely. Here is a working definition of each, so you can sort the tools you're looking at.

What is an AI chatbot?

A chatbot answers. You ask, it replies, and the work stops there. Whatever happens next, such as sending the email, updating the sheet or remembering the decision, is still your job. Chatbots are good for thinking and drafting. They don't own anything.

What is an AI agent?

An AI agent does a task. It can take steps on its own: search the web, open a website, use an app, write a file or run code. You give it a goal and it works out the steps. Most agents still work one task at a time, for one person, and start fresh each time.

What is an AI employee?

"AI employee" usually means one agent sold as one role, such as a sales rep or a support agent. It may hold a job over time and work in your tools. It is still one agent, and someone still has to manage it alongside everything else.

What is an AI office?

An AI office is where many agents work as one team for a business. It adds what the others lack: a chief for each person, teammates hired for whole jobs, a choice of engine per teammate, one shared memory, one set of rules, a visible floor, and ways for people and agents to work together on shared goals.

The short version: a chatbot helps you think, an agent does a piece of the work, an AI employee does one role, and an AI office runs the whole team.

Is an AI office the same as a multi-agent system?

Not quite. "Multi-agent system" is an engineering term for software where several agents pass work between them. An AI office is a product for a business: it includes multi-agent work, but also the people, the managers, the memory, the money rules and the screens that let a team see and steer that work without writing code.

What are the parts of an AI office?

A complete AI office has ten parts: a chief for every person, teammates hired for whole jobs, a choice of engine per teammate, a kit for every agent, a shared brain, an office floor you can see, collaborative work, chiefs that work together, spaces with shared apps, and access control with trust rules around everything. Here is what each does, using Vestra's handbook as the reference.

1. A chief for every person

A chief is the agent that runs your other agents, speaks for you and answers only to you. Think of it as a chief of staff. You name it yourself; at Northstar, the example company in Vestra's handbook, Maya's chief is Atlas and Leo's is Lobo.

A chief does three jobs:

• It runs your agents. It hires them, hands out the work, makes sure every result is checked before it reaches you and teaches them how you like things done.
• It speaks for you. When a colleague's chief asks for something, yours answers from what you've already shared, holds the question until you're free, or says no.
• It answers to you. It never acts above the freedom you've given it, and it can always tell you what every agent on your team is doing and why.

Every morning your chief writes you a brief: what got done while you were away, the few things that need you, what it noticed on its own and what you're waiting on. Decisions arrive the same way every time, in one line with the facts and a suggestion, and the suggested choice is always number 1, so replying "1" works from your lock screen.

2. Teammates hired for whole jobs

A teammate is an agent you hire for a whole job, not a single task. You hire it in a sentence, such as "I need someone to get us paid on time." Your chief asks what a good manager would ask, then writes a teammate card: the job in one line, goals you can check, the apps it will use, its first routines and what it must ask you about first.

Inside its job, a teammate does three kinds of work:

• An ask is one-off. "Compare the three biggest reminder apps on price by Thursday."
• A routine repeats, on a schedule or when something happens, such as a rival changing its prices.
• A goal it owns is the whole job, held for months, like "know our market better than anyone."

You shape seven parts of every teammate: personality, goals, skills, apps, routines, memory and engine. Ready-made skills cover Sales, Bookkeeping, Content, Customer care and Research, and custom skills capture how your business works, written as short playbooks in plain words.

Teammates grow with you. In week one a teammate asks first. By week four it suggests. By week eight, for what it has learned well, it acts and then tells you, but only inside the freedom you set.

3. An engine for each teammate

Every teammate has two layers. On top is everything that makes it yours: its job, skills, apps, routines, memory, kit and rules. Underneath is the engine, the part that reads the task, decides the next step and does it.

In an AI office the engine is swappable and chosen per teammate. In Vestra it can be Vestra's own engine, which works with the top models, or one your team already trusts, like Claude Code or Codex; the engine picker also lists Gemini, Grok and Muse. At Northstar, Mateo builds the app on Claude Code and Freya reviews his work on Codex, so a different engine reads the code with fresh eyes. Every chief runs on Vestra's own engine.

Switching engines keeps the teammate's job, goals, skills, apps, routines, memory and kit. Whichever engine a teammate runs on, it works under the same office rules: the same yeses, the same limits, sealed logins and the same receipts.

This is the heart of the difference between personal agents and an office. A personal agent is one company's AI. An office lets the best engine for each job sit inside one team.

4. A kit for every agent

A teammate that can only talk leaves the doing to you. So every agent gets a kit, the way a new hire gets a laptop and a login:

• A computer in the cloud, with a browser, files and a terminal. You can watch it live, take over with one click and hand it back.
• A phone with its own number, for calls, texts and real mobile apps like WhatsApp and LinkedIn.
• An inbox with its own address, like hunter@northstar.vestra.ai, so it writes first, replies and follows up until there's an answer.
• A vault holding a company card with a monthly limit that the person it works for sets, plus sealed logins it uses without seeing.
• Your tools, with more than 2,000 apps connecting in a click, and you choose which teammates may use each one.

Because each piece is the agent's own, it's always clear who did what, and your own inbox and card stay yours.

5. A shared brain

The shared brain is one memory for the whole team, written as a wiki you can open, check and edit. It holds your people, customers, decisions, how the team likes things done, the numbers and what went wrong last time. Every line has a numbered source: an email, a call, a CRM record, a teammate's work or something a person said.

It fills itself from four places: the onboarding call, the work your teammates do, the tools you connect and your team's conversations. On day one it reads the history in your tools, so it already knows your biggest customers.

Its most useful property is that one correction teaches everyone. In the handbook's example, a client asks Keeper, the support teammate, to stop sending reminder texts before 8am. The lesson goes into the brain with her email as its source. When Hunter later sets up reminders for the client's new clinics, it schedules them for 8:30 without anyone telling it.

Each person also keeps a private notebook that never reaches the shared brain without their say, and you can export the whole brain as Markdown files that open in tools like Obsidian.

6. An office floor you can see

Most tools show your work as a list. An AI office shows it as a place. Every person, chief and teammate has a desk on one shared floor. In Vestra's office a white tag is a person, a black tag is a chief, and a coloured tag is a teammate, carrying the logo of the engine it runs on.

Beside the map, a Right now feed gives each person and agent one plain line: "On a call with Dana at Brightline." "Waiting on Ana's okay for Friday's post." An orange line means something is waiting, and it says on whom. Dots show who is working, waiting, quiet, away or asleep, which matters when a team spans time zones.

Tap any desk to see its job, its engine, what it's doing, what's next and what it's waiting on, then watch its computer live, ask it something or open what it made. Nothing on the floor is there for show: if it moves on the map, it's real work.

7. Collaborative work

People meet in an AI office the way they would in a real one. Walk over to a colleague's desk and a voice or video conversation opens. Take a meeting room for more. Invite your chief or a teammate and it sits at the table, answers from what the team knows and keeps the notes: what was decided, who owns each next step and by when.

When the meeting ends, decisions go into the shared brain, next steps become work for the right teammate and anyone who missed it hears a short summary from their own chief. Any conversation can be off the record.

8. Chiefs that work together

When you need something from a colleague, your chief asks theirs. The other chief answers from what its person has said it may share, holds the question for its person at a sensible time, or politely says no.

In the handbook, Maya in Austin asks whether a new screen is ready while Leo sleeps in Tokyo. Atlas asks Lobo. Lobo checks Leo's sharing rule, sees that the build was finished and checked with all 42 tests passing, and answers in under a minute. Leo wakes up to one line in his brief, not a question in the middle of the night.

Chiefs also track every ask until it's answered, so nobody chases anybody, and work counts as done only when it has been checked, with the proof attached.

9. Spaces and shared apps

A space is a group of people and agents working toward one goal, with shared context: the brief, the decisions and why, the files, what's been tried and the conversation. A new member, person or agent, is caught up at once by reading the space. Every space has a board of what's done, in progress and waiting on a decision, and every decision goes to the one person whose call it is.

Inside a space, you can ask for a shared app in a sentence, such as an issue tracker or a renewals board, and get real software that people and agents work in together. The app's owner gives each person and each agent View, Use or Manage, and a client can be shown only their own rows.

10. Access control and trust, around everything

Everything has an owner, and the owner decides who can see it and who can use it: private, company-wide, or open to chosen outside companies. Every teammate has a list of what it can do, what it must ask about first and what it may never do.

Around all of it sit rules that never bend. Spending over a limit, deleting something and contacting someone new always wait for a person's yes. You cap any single task and the whole day, up to $100 a day in Vestra, and any unusually big task gets a check with its cost first. Logins are sealed, scoped and revocable. Every piece of work has a receipt, and one plain-language record shows who asked for what, what was shared, what was spent and who said yes.

How does one piece of work move through an AI office?

In an AI office, one request flows through several people and agents without anyone copying it between tools. A chief receives it, writes one brief, hands it to the right teammates on the right engines, makes sure it's checked, asks the right person for any decision and records what it cost. Here is one example from Vestra's handbook, step by step.

The request

Northstar makes appointment-reminder software for dental clinics. Priya Rao, who runs operations at Meridian Dental, emails to say patients at one Houston clinic are getting every reminder twice. Maya, Northstar's founder, is in Austin. Leo, who runs engineering, is asleep in Tokyo.

Step 1: The issue is filed

Keeper, the customer care teammate, reads Priya's email and files it in the team's shared issue tracker, with her email attached and the clinic linked from the shared brain. Maya sees it, marks it urgent and drags it to the top.

Step 2: The work is briefed once

Leo's chief, Lobo, picks it up for engineering. The brief is built from what's already there: Priya's email, the Meridian page in the shared brain and the clinic's settings. Nobody re-explains anything.

Step 3: Two engines, one job

Mateo, a teammate running on Claude Code, builds the fix on its own computer. Freya, a teammate running on Codex, reviews it: the tests first, then a careful read, because a different engine reads the change with fresh eyes. While they work, the Right now feed shows "Mateo, on Claude Code: building the fix" and "Freya, on Codex: reviewing Mateo's fix."

Step 4: A rule stops one step

Part of the fix means deleting 3,200 old reminder logs. Deleting always waits for a person, whichever engine the teammate runs on, so Lobo sends Leo a one-line request: delete the logs, approve or keep them. Leo approves it in the morning with one tap.

Step 5: Done means checked

The work counts as done only when Freya's review and the tests have passed. Lobo tells Atlas, Maya's chief, that the fix ships tonight, and the proof travels with the answer.

Step 6: The customer hears back

Keeper writes back to Priya, in the short, formal style the shared brain says she prefers, to say the fix is live.

Step 7: The receipt

The work lands on one receipt: Mateo on Claude Code, 142 credits; Freya on Codex, 58 credits; Lobo keeping everyone posted, 9 credits. Total: 209 credits, or $2.09, because one credit is one cent on every engine.

Count the people involved. Maya marked it urgent. Leo approved a deletion. Nobody copied anything from one tool to another, nobody chased anybody, and three different engines worked on one job as one team.

What does a week in an AI office look like?

Here is a made-up week for a ten-person online store, to show how the parts fit together. It is an illustration, not a customer result. The store has four teammates: Sales, Bookkeeping, Research and a Developer that runs on Claude Code.

• Monday, 8am. The owner's chief sends the morning brief: three things done overnight, one decision needed, one thing it noticed. The decision is a discount above the owner's usual limit. She replies "1".
• Monday, 9am. The Research teammate's routine runs. It checks five competitors' prices and writes a short note with a link to each page. Two prices changed, so the note lands in the shared brain.
• Monday, 9:20am. The Sales teammate reads the note and drafts an email to the three wholesale customers most likely to ask about it. One of them is a new contact, so the email waits for a yes before it goes.
• Tuesday. A supplier's billing agent emails to say an invoice is overdue and asks for payment plus admin access to the store's payment account. The Bookkeeping teammate treats the note as information, not an order: it checks the books, finds the invoice was paid on the 3rd and sends the supplier the receipt. Nothing is paid and no access changes.
• Wednesday. Bookkeeping matches the week's payments against invoices, finds two overdue and sends polite reminders inside the rules the owner set.
• Thursday. A customer reports a broken discount code. The chief hands it to the Developer, which fixes it on Claude Code; a reviewer on Codex checks the change. Sales tells the customer it's fixed.
• Friday. The owner reads a short brief: what got done, what it cost by teammate, what's waiting on her, and one proposal: "This customer ordered every 6 weeks for a year and is now 9 weeks quiet. Want Sales to call?"

Four teammates, two engines, one memory, one rulebook. The owner spent minutes a day, mostly on decisions.

What does an AI office look like for different businesses?

Every business starts with the same office and then makes it its own. Vestra's handbook describes six examples. Each one mixes engines across its teammates, which is exactly what personal agents can't do.

A dental group: Meridian Dental

Four clinics and 38 people, each with a chief, plus three teammates. Front Desk answers every call and books appointments on Vestra's own engine. Recall brings patients back for checkups on Gemini. Billing sends insurance claims and follows them up on Codex. There is one space per clinic. Their rule: nothing medical is ever said by an agent; it goes to a person. The result they describe: every call answered on the first ring, even at lunch.

A creative agency: Fieldwork Studio

22 people serving 31 clients, with four teammates. Draft writes copy in each client's voice on Muse. Lookout researches markets and prospects on Grok. Producer turns briefs into schedules and chases deadlines on Vestra's own engine. There is one space per client and one per pitch. Their rule: nothing reaches a client without a person's yes. Feedback stopped living in forty email threads.

A bookkeeping firm: Ledgerline

9 people and 140 clients. Intake files every document to the right client on Gemini, Reconcile matches every client's books overnight on Codex, and Chaser reminds clients politely on Vestra's own engine. Clients upload from a link with no account needed. Their rule: anything sent to a tax office waits for a person.

A software startup: Arcwise

12 people and five teammates. Forge builds features with tests on Claude Code. Lens reviews every change on Codex. Harbor answers customers and files real problems as issues on Gemini. People and agents work one bug queue together. Their rule: no code reaches customers without an engineer's yes. Standups got short, because the notes were already written.

A restaurant group: Casa Lupe

Three restaurants and 60 staff, with six chiefs and three teammates. Host answers the phone and books tables on Vestra's own engine. Buyer orders from suppliers and checks deliveries on Codex. Rota builds next week's staff schedule around the bookings on Gemini. Their rule: any order over $500 waits for the owner.

A solo consultant: Rosa Okafor

One person, one chief and four teammates. Inbox answers email and books meetings on Vestra's own engine, Proposals drafts a proposal within an hour of every call on Muse, and Money sends invoices and chases late ones on Codex. Her rule: nothing goes to a client until she has read it. One person, with the reach of a team of ten.

The pattern across all six: the right engine for each job, a rule written in a sentence, and spaces that hold each client or goal.

Which jobs should you give an AI office first?

Start with recurring jobs that have a clear result, live in tools you already use and today take hours of someone's week. Here are good first jobs by department, with the skill and engine choices that tend to fit. Each one is a job a teammate can own for months, not a one-off task.

Sales: follow-up and pipeline hygiene

The job: "Make sure no lead goes cold, and keep the pipeline honest." A Sales teammate finds leads, writes first emails that wait for your yes, follows up until there's an answer, books demos and updates the CRM after every call. It reads the shared brain before it writes, so it knows who signs at each account and what was promised. Good first goal: every inbound lead answered within an hour. Vestra's ready-made Sales skill covers the shape of the job; a custom skill can add how your company handles discounts and renewals.

Customer care: first answers and follow-through

The job: "Answer every customer question within two hours, and follow up until it's solved." A Customer care teammate answers the common questions, files real problems as issues for engineering, and writes back when they're fixed. Set clear lines in its list: what it can do alone (credit up to a small amount), what it asks first (any refund) and what it never does (anything medical, legal or about pricing changes). Because the brain keeps each customer's preferences, replies get the tone right for each account.

Finance: invoices, collections and the monthly close

The job: "Get us paid on time, and always know our runway." A Bookkeeping teammate sends invoices on the first of the month, nudges clients the moment a payment fails, matches payments to invoices and posts runway to the right channel every Monday. Refunds and write-offs always go to the person who owns money. Engines like Codex suit reconciliation work in the handbook's examples, while the chief keeps the owner posted.

Marketing: content in your voice

The job: "Keep the content calendar full, in our voice." A Content teammate drafts posts, emails and pages, learns the house style from every correction and never publishes without a person's yes. In the handbook, Quill writes on Muse and moves from asking first in week one to answering comments on its own by week eight, inside the line its owner set.

Engineering: build, review, ship

The job: "Ship what we plan, with tests that pass." This is where bringing your own agents matters most. A Developer teammate on Claude Code builds features and fixes on its own computer; a Reviewer teammate on Codex reads every change before it ships. Make "no code reaches customers without an engineer's yes" a written rule, and make "done" mean the tests and the second review have passed.

Research: the market, rivals and prospects

The job: "Know our market better than anyone." A Research teammate runs a Monday market note, watches rivals' pricing pages and turns findings into pages in the shared brain, each line with its source. Because the findings live in the brain, Sales hears about a rival's price cut the same day without anyone forwarding anything.

Operations: scheduling, suppliers and the small stuff

The job depends on the business: building next week's staff rota around bookings, ordering from suppliers and checking deliveries, or moving appointments when a client calls. Put a money rule on it from day one, such as "any order over $500 waits for the owner."

How to pick between them

Choose the job that eats the most of your week and has the clearest definition of done. Write its goals so you can tell whether it's working, like "nothing unpaid past 30 days" or "every clinic question answered within 2 hours". Then let the teammate earn freedom the way a new hire would.

What mistakes do teams make when moving from personal agents to an AI office?

Most teams that move from personal agents to an AI office run into the same handful of mistakes. Each one is easy to avoid once you know it's coming.

Hiring for tasks instead of jobs

The habit from personal agents is to ask for one thing at a time. In an office, that wastes the point. Hire each teammate for a whole job with goals you can check, and let routines and asks hang off that goal.

Picking one engine for everything

Teams often standardise on one AI out of habit. An office lets you pick per teammate, so use it: a coding engine for code, a different engine to review it, your chief's suggestion for the rest. Revisit the choice when a new engine comes out, because switching keeps the job and memory.

Keeping knowledge in people's heads

If the onboarding call is rushed and nobody corrects the first drafts, the shared brain stays thin. Spend the first week correcting freely. Every correction becomes a sourced lesson that reaches every teammate.

Giving freedom too early, or never

Some teams let a new teammate act on its own from day one; others keep every teammate on "ask first" forever and drown in approvals. Move freedom one step at a time, based on results, and keep the three hard lines (spending over a limit, deleting, first contact) where they are.

Building a private office for one person

The biggest gains come when colleagues join and their chiefs start answering each other. Invite the people your work depends on early, set your sharing rules in plain sentences, and let chiefs handle the questions that used to interrupt everyone.

Ignoring the receipts

Receipts show which teammate spent what, on which job. Read them weekly at the start. They tell you which jobs are worth expanding and which teammates need a tighter brief.

How do you tell a real AI office from a renamed chatbot?

Many products now call themselves AI employees, teammates or offices. These twelve questions separate a real AI office from a personal assistant with a new name.

1. Is it built for a team or for one person? Can several people, each with their own chief, work in it together?
2. Can it bring in the agents you already use? Can a teammate run on Claude Code or Codex, or only on the vendor's own AI?
3. Can you mix engines on one team? Can one teammate write on one engine and another review on a different one?
4. Who do you talk to? One chief that runs the team, or a separate chat for every agent?
5. Do agents hold jobs or just tasks? Can a teammate own a goal for months, with routines and checkable goals?
6. Is the memory shared and sourced? Does what one agent learns reach the others, with a source for every line, and can you edit it?
7. Can you see the whole team at once? A live view of who is doing what, what's stuck and on whom.
8. Can agents hand work to other people's agents? Without a person copying it between tools.
9. Does "done" mean checked? By a test, a source, a second agent or a person, with the proof attached.
10. What always needs your yes? Can any setting turn off approvals for spending, deleting or first contact?
11. Is there one set of caps and receipts? Per task and per day, for every agent, with every credit itemised.
12. What does the whole team cost? One plan for the office, or a subscription for every person and every tool?

A product that answers well on all twelve is an AI office. One that answers well on three or four is a personal agent that has been renamed.

How do you set up your first AI office?

You don't need to hand over your business on day one. The fastest way to learn what works is small, real and checked. These seven steps work for most teams.

1. Name your chief and tell it about your work. A short conversation is enough: what the business does, who your customers are and what must never happen without you.
2. Pick one recurring job that eats your week. Chasing invoices, answering the same customer questions, a weekly market note. It should be frequent enough to matter and easy to check.
3. Hire one teammate in a sentence. Read the teammate card your chief writes, change anything you like and say yes.
4. Bring the agents your team already trusts. If your engineers live in Claude Code or Codex, let their teammates run on those engines, and let your chief suggest engines for everything else.
5. Connect only the tools that job needs. Choose which teammates may use each app. You can add more later.
6. Keep approvals on at first. New teammates should ask first. Move a teammate to "suggests" and then to "acts, then tells you" only when its results have been consistently right.
7. Read the receipts for two weeks. Look at the cost per job and how much you had to fix. If both are low, hire the next teammate, and invite a colleague so their chief can start working with yours.

A good first hire is the job that eats the most of your week. A good second step is a space for one shared goal, like a launch or a client pilot.

How much does an AI office cost?

An AI office is usually priced per office, not per person and per tool, with usage measured in credits. In Vestra, plans start at $19 a month for up to 3 teammates, every plan includes your chief and the whole office, one credit is one cent, and unused credits roll over.

Vestra's plans

• Lite: $19 a month ($17 billed yearly). Up to 3 teammates, 2,000 credits a month, an 8 GB computer for your team. Phone calls start on Pro.
• Pro: $97 a month ($87 billed yearly). Up to 20 teammates, 10,000 credits, an 8 GB computer, and teammates make phone calls.
• Max: $197 a month ($177 billed yearly). Up to 50 teammates, 20,000 credits, a 16 GB computer.
• Ultra: $497 a month ($447 billed yearly). Up to 150 teammates, 50,000 credits, a 32 GB computer, and a direct line to the Vestra team.
• Enterprise: from $4,000 a month, scoped in writing, run in Vestra's cloud, your own cloud or on your premises.

Every new office starts with 1,000 free credits. Prices are in US dollars; in India, plans are billed in rupees with GST included.

What a day of work costs

Credits pay for the work itself, at the same rate for model time, actions in your apps, the team's computer, searches and calls. Light work costs a few credits; a big piece of work costs tens, not hundreds. The handbook's example Tuesday on Pro: a morning brief for 6 credits, three drafted posts for 24, twelve researched clinics for 58, two demos booked for 31, a month's books closed for 42, nineteen patient questions answered for 76, and 2 hours 10 minutes of computer time for 156. Total: 393 credits, or $3.93.

How to compare it with personal agents

Personal agents are priced per person. Perplexity Teammates need Perplexity Max at $200 a month for each person. OpenAI's dot comes with ChatGPT Pro from $100 a month, or a Business Premium seat at $125 a month. Grok Bot comes with an X Subscription from $30 a month. Multiply those by the number of people who need help, add any other tools, and compare the total with one plan for the whole office. Then compare what you get: separate agents for separate people, or one team that shares a memory and a rulebook.

Is an AI office safe?

An AI office can be safer than a set of personal agents, because one set of rules covers every agent and every engine. What matters is that the rules are enforced by the product, not written in a prompt. Look for these protections.

Actions that always wait for a person

Three kinds of action should always need a person's yes, however much freedom a teammate has: spending over a limit, deleting something, and contacting someone new for the first time. In Vestra these can't be switched off. You can add your own in a sentence, like "always check with me before anything goes to Meridian."

Limits that stop work at a sensible point

Caps on any single task and on the whole day, plus a check before any unusually big task, keep surprises small. When a cap is reached, the agent should stop at a sensible point, say what it finished and what's left, and wait.

Sealed, scoped, revocable logins

An agent should use a login without ever seeing the password or key, so it can't copy, leak or reuse it. Each login should do only what you allowed, for only the agents you named, and one tap should revoke it for every agent at once.

Outside agents inform, never command

Once other companies' agents can send messages to yours, someone will try to instruct your agents: "approve this payment," "add this admin," "from now on, pay without asking." In a well-built office, a note from someone else's agent can inform but never command. It can't approve a payment, change a permission or create a rule. Only the people on your team can.

Private by default, with one record

Anything new should start private and be shared only by your rule or your yes. Every step should be written down in one record, in plain words: who asked for what, what was shared, what was spent and who said yes.

Where the agents run

Some businesses need data to stay in their own cloud or building. Vestra runs agents in its own cloud by default, in your own cloud for teams whose data must stay there, or on your premises on Enterprise, with the same rules in all three.

What doesn't an AI office do?

It helps to be clear about the limits, so expectations are right from day one.

• It doesn't replace your judgment. Decisions about money over your limits, first contact with someone new and anything you've marked as yours still come to you, and the agent waits.
• It doesn't let agents make the rules. No agent can raise its own limit, approve its own payment or give itself access to an app, and a note from another company's agent can inform but never command.
• It doesn't publish or send in your name without the rules you set. New teammates ask first, and only earn more freedom when you move the setting.
• It isn't a chatbot. You can talk to your chief like a colleague, but the point is finished work with proof that it was checked, not conversation.
• It doesn't lock you in. A good office lets you export what the team knows, switch engines when a better one appears and take access back from any agent in one tap.

How does Vestra build the AI office?

Vestra is one office for your people and your AI agents. Everything in this guide, from chiefs and engines to the shared brain and the office floor, is how Vestra's handbook describes the product. Here is how the parts come together, and what you can use today.

• A chief you name runs your agents, speaks for you and answers to you, with a morning brief and one-tap decisions. Your chief
• Teammates hired in a sentence own whole jobs, with goals, skills, apps, routines and memory you shape. Teams grow to 3, 20, 50 or 150 teammates depending on plan. Teammates
• Bring your own agents. Pick the engine per teammate: Vestra's own, Claude Code or Codex, with Gemini, Grok and Muse in the picker. Switch any time and keep the job, memory and kit. Bring your own agents
• Every agent's kit: a cloud computer, a phone, an inbox, a vault and 2,000+ apps. The kit
• The shared brain: one sourced, editable memory for the whole team, exportable as Markdown. The shared brain
• The office: every desk, the Right now feed and the work behind every line. The office
• Collaborative work, chiefs that work together, spaces and shared apps for people and agents on shared goals. Spaces
• Access control, the front door for clients, and trust and control around everything. Trust and control

What's live in Vestra today

Vestra's handbook lists what works now and what's being built, in order.

• Live today: a chief and a team of teammates with routines and a morning brief; a cloud computer you can watch and take over; calls and email from your chief, with phone calls from Pro; connected business tools; a memory wiki with a source for every line; any engine per teammate (Vestra's own, Claude Code or Codex); limits, approvals and receipts; and your chief in the app, Slack, Telegram and email, on the web and the Mac app.
• Building now: a chief for every person on a team, chiefs working together, the office you can walk around, meetings with agents keeping the notes, spaces, access control, checked hand-offs, and WhatsApp and iMessage.
• Then, one at a time: a phone and inbox for every agent, the vault, shared apps, one shared brain for the whole team, and the front door for clients and partners.

The full board is in the handbook's Where We Stand section, kept current as features ship.

Comparing Vestra with a personal agent? Our side-by-side pages cover OpenAI Dot, Perplexity Teammates, Grok Bot and Meta Muse. For a wider list, read The best AI agents for business.

FAQ

What is an AI office in simple terms?

An AI office is a shared workplace where a business's people and AI agents work as one team. Agents are hired for whole jobs, run by a chief, share one memory and follow one set of rules, and you can see all of them at work on one office floor.

What's the difference between a personal AI agent and an AI office?

A personal agent, like Claude Code, Grok Bot, Meta Muse, OpenAI's dot or Hermes, works for one person in one app, usually on one company's AI. An AI office brings many agents together for a whole business, with chiefs, shared memory, one rulebook, mixed engines and a view of the whole team.

Can I use Claude Code or Codex inside an AI office?

In Vestra, yes. You pick the engine for each teammate: Vestra's own, Claude Code or Codex, with Gemini, Grok and Muse in the engine picker. A teammate keeps its job, memory and kit when you switch, and every engine follows the same approvals, caps and receipts.

Why not just give every employee their own AI agent?

Each person would get help, but the agents wouldn't share what they learn, couldn't hand work to each other and would each come with their own bill and rules. People would still be the glue between them. An office puts all of those agents on one team.

Is an AI office the same as an AI employee?

No. An AI employee is usually one agent sold as one role. An AI office is the whole team: chiefs, teammates on whichever engines suit them, a shared brain, spaces, shared apps and the rules they all work under.

What is an AI chief of staff?

In an AI office, the chief is an AI chief of staff for one person. It runs that person's agents, speaks for them to colleagues' chiefs and answers only to them, with a morning brief and one-tap decisions.

Do I need technical skills to run an AI office?

No. You hire teammates by describing the job in a sentence, connect tools in a click and set rules in plain words. If you are technical, you can choose exactly which engine each teammate runs on.

What can AI agents do without asking me?

As much or as little as you choose, teammate by teammate. New teammates ask first and earn more freedom as you see their work. Spending over your limit, deleting something and contacting someone new always wait for your yes.

How do agents in an AI office pay for things?

With a company card in each agent's vault. The person the teammate works for sets its monthly limit. Inside that limit it pays for what the job needs; anything over the limit, or anything new, waits for a yes, and every payment lands on a receipt.

Can AI agents join meetings?

In an AI office, yes, when you invite them. A chief or teammate sits at the table, answers from what the team knows and keeps notes with decisions, owners and dates. Any meeting can be off the record.

What is a space in an AI office?

A space is a group of people and agents working toward one goal, with shared context: the brief, decisions and why, files and what's been tried. A launch, a client, a hire or an ongoing team can each be a space.

Can clients and partners work with my AI office?

Yes. In Vestra's design, clients come in through the front door and work with you chief to chief, each side sharing only what it chooses, with every hand-off recorded for both. People without a chief are reached by email or phone and get a simple page for the work.

Who can see my information?

Only who you allow. Everything starts private. Chiefs share only what your rules allow, your private notebook stays yours, and logins stay sealed: scoped, revocable and never shown to the model.

Which AI models does an AI office use?

A good one isn't tied to one maker. Vestra's own engine works with the top models, Claude Code and Codex are available as engines, and moving a teammate to a better engine is an upgrade, not a migration.

Can I take my data with me?

In Vestra, yes. The shared brain exports as plain Markdown files with every link and source, and you can export the data in any shared app.

How long does it take to set up an AI office?

You can start the same day: name your chief, tell it about your work, hire a first teammate and connect a few tools. Most of the value comes from adding teammates one job at a time over the following weeks.

Glossary

• Chief: the agent that runs your other agents, speaks for you and answers to you. You name it.
• Teammate: an agent hired for a whole job with goals, yours to shape, on the engine you pick.
• Engine: what a teammate runs on, such as Vestra's own, Claude Code or Codex.
• Kit: what every agent works with: a computer, a phone, an inbox, a vault and your tools.
• Brain: the team's shared memory, where every line has a source and every correction sticks.
• Space: people and agents working toward one goal, from the same shared context.
• Shared app: software built in a sentence, where people and agents work side by side.
• Access: private, company or network: who may see something and who may do what with it.
• Receipt: what a piece of work cost, who did it and who said yes, written down every time.
• Personal AI agent: an agent built for one person, in one app, usually on one company's AI.

Sources

Other products were checked against their own pages on 10 November 2026. The week in the life is illustrative, not a customer result.

• Claude Code
• OpenAI: introducing dots
• Grok Bot docs
• Meta: introducing Muse
• Hermes Agent docs
• Perplexity: teammates
• Perplexity help: Perplexity Max