The operating system for digital employees.
Most companies have employees and software. Soon they will have AI workers too. AgentOS gives each one an identity, a memory, clear permissions and real tools, runs it for you, and asks before anything important happens.
We are letting people in a few at a time.
From a sentence to a working digital employee
No flowcharts and no code. You say what you want done, check what the worker will do, and put it to work.
Describe the job
"Each morning, summarize anything urgent in my inbox and list today's meetings."
Review and set the limits
AgentOS drafts the worker: its goal, its tools and whether it may send or change anything. You edit it and set its permissions before it starts.
Deploy it
Run it on demand, on a schedule, or from a web address your other systems can call. Every run is recorded with its cost.
Six layers every digital employee needs
Own these layers and every specialist AI worker can run safely on one platform, instead of living in a separate chat window.
1. Identity
Who is this agent?
Like a person on your team, every agent has a name, a job title, a handle and a home workspace. You can see who created it and who manages it, switch it off in one click, or give your own software a token that acts as that agent and can be revoked at any time.
2. Memory
What does it remember?
Each agent keeps its own notes between runs: customer preferences, past results, lessons learned. The workspace keeps shared business knowledge such as policies, pricing rules and house style. Workers get more useful the longer they work.
Memory cannot be used to plant standing orders: text that tries to instruct the AI is refused.
3. Permissions
What is it allowed to do?
Not "everything". Limit the tools it can use, the domains it may email, the Slack channels it may post to, how many times it runs a day and how much it may spend. Anything that sends or changes something waits for a person.
4. Connectors
What can it work with?
Without connectors an agent is a chatbot. With them it is a worker. You connect your own accounts and choose exactly which permissions to give.
5. Marketplace
How do people find and install agents?
Browse ready-made workers, read exactly what they do, and install one in minutes. Installed agents start read-only, so you decide what to turn on. Every listing is reviewed first, can be rated, and can be reported.
Builders can publish for free today. Paid listings with a revenue share for the builder are coming.
6. Deployment
Where does it run?
You do not manage servers. Create, deploy, work. Freeze a version so edits never change a live worker, run it hourly, daily or weekly, or call its web address when something happens. Every run, cost and approval is logged, and one switch pauses everything.
How the layers work together
Take a recruiting coordinator for a growing company.
Digital employees you can put to work
Each of these uses the tools that work today: Gmail, Google Calendar, Slack, Notion and HubSpot.
Lead follow-up for home services
A new enquiry arrives from your web form. The worker drafts a friendly reply, proposes a time on the calendar and remembers the customer's preferences for next time.
For plumbers, HVAC, cleaners, landscapers.
Home care scheduler
Reads caregiver availability emails, drafts shift offers for your approval, and posts tomorrow's roster to Slack.
Never confirms a shift without a person.
Recruiting coordinator
Reads new applications in Gmail, proposes interview slots from the calendar and posts a daily shortlist to Slack.
Cannot make an offer.
Support triage
Reads the customer's email, finds the answer in your Notion help pages, drafts the reply and flags urgent cases in Slack.
Sales researcher
Before each call, looks up the contact in HubSpot, reads past email and prepares a one-page brief.
Finance assistant
Reads invoice emails and drafts polite payment reminders.
Can read and draft. Cannot send without approval.
Chief of staff
Every morning: the urgent mail, today's meetings and what changed in your Notion workspace, in one short note.
Agency account manager
Answers client status questions from your project notes and drafts the reply for the account lead to approve.
It asks before it acts
An agent that reads your email should never surprise you. These protections are on by default.
Agent finished: "I drafted a reply to Sue about Thursday's meeting."
Waiting for your approval: send emailTo: sue@example.com · Subject: Re: Thursday
Approve to send, or reject and nothing happens.
Approval before every send
Emails, calendar changes and Slack posts wait for a person. Approving does exactly what you saw, once.
Practice runs
Dry runs read your real data but only preview what the agent would send.
Warnings that matter
If an agent is about to write to someone you never mentioned, the approval says so.
Defends against tricks
Emails can hide orders for an AI. AgentOS treats what it reads as data, flags suspicious text, and blocks sending and saving memories for that run.
One-click pause
A single switch stops every run, schedule and approval for your workspace.
Full audit trail
Who connected what, which agent ran, what was approved. The log cannot be edited.
Connect the tools you already use
You connect your own accounts and choose exactly which permissions to give. Tokens and keys are encrypted, and you can disconnect at any time.
Working now
In beta
Planned
A workforce with three kinds of members
The companies that do well will manage people, AI workers and software together. AgentOS is the system that manages the AI workers.
Questions and answers
Put your first digital employee to work
Tell us where to send your invitation.