AI employees
Give an AI a goal, and it finds work to do independently.
Set goals in a meeting, steer it through the Inbox, and handle requests when they arrive.
Understanding AI employees
An AI employee works independently toward a goal you assign. It does not follow a fixed sequence like a package. Each shift, it reads its goal and past memory, then chooses the work most needed now. It also finds new places and methods suited to the goal, beyond the ones you suggest.
You and the employee have different responsibilities.
| What | Who handles it | Example |
|---|---|---|
| Goal | The AI drafts it in a meeting; you confirm it | The goal statement, things not to do, and facts to know. Numbers and deadlines, such as “500 subscribers within three months,” are included only when you specify them. |
| How to get there | The employee independently | Where, what, in what order, and how often to work; new places to try; to-do lists; lessons learned |
| Work conditions | You (in Work settings) | Working AI · Model · Reasoning effort · Work schedule · Work hours |
| Sign-up and sign-in | The employee | Uses accounts already signed in to the orke browser (Google, GitHub, etc.) and buttons such as “Continue with Google.” It does not enter or create passwords. |
| Passwords, verification, and payment | You | Entering passwords; phone, SMS, email, and identity verification; payment; CAPTCHAs (tests that verify you are human). The employee sends a request. |
| Recurring tasks (packages) | AI employees | |
|---|---|---|
| How they work | Run the saved sequence as written | Choose their own tasks based on the goal |
| Memory | Start fresh with each run | Continue through status, plan, lessons, and journal files |
| Communication with you | None during execution | You send instructions through the Inbox; employees request only things a person must do |
| Good fit | Tasks with a clear sequence | Ongoing work, such as research, promotion, and organization |
There are no daily or shift reports. Employees find and decide on materials, information, numbers, permission, and direction themselves. They leave requests only when something requires a person, such as a password, verification, or payment, and do other work while waiting. You can check their work in the Journal at any time.
What to prepare before starting
- Working AI — Install and sign in to Claude Code, Codex (ChatGPT), NVIDIA, or Antigravity. See AI setup and accounts.
- Sign in to the orke browser — Employees use the same browser profile as orke web windows. If you sign in to an account such as Google or GitHub in an orke web window, employees use buttons such as “Continue with Google” to sign up and sign in to the sites they need. For password-based sites, sign in beforehand in an orke web window so the employee can work with that signed-in state.
- Keep orke running — Employees work only while orke is running. You can close the window and leave it in the tray. When the PC is off, work pauses too.
The first example below organizes research in the employee’s memory files. Start with “Do not publish anything anywhere” in the goal document’s “Things not to do” section so you can learn the workflow without publishing externally.
Example: create your first AI employee
Create a “research employee” that collects and organizes daily news about open-source projects you follow.
Open New AI employee
Press “Packages” in the Dock to open the Packages window. Press “+” in the “AI employees” group header to open “New AI employee.” Describe the assignment in one or two lines under “What should it take on?”, then select the “AI to work with.”
Press “Start meeting” to create the employee folder and open a meeting. You will choose a name in the meeting.
✓A “New AI employee” row appears in the “AI employees” group, and “Meeting with the AI” opens on the right.
Set the goal in a meeting with AI
The AI in the left terminal reads the assignment and, if needed, checks public facts such as your homepage. It fills out the entire “Goal document” on the right first. Then it asks just once: “Shall we start with this? Let me know if anything needs changing.” Tell it what to change, or say it looks good. Numbers and deadlines are added only when you specify them.
When a new release comes out, summarize the changes in three lines.
Let’s aim for 10 release summaries by the end of this month.
Check that “Things not to do” includes a line saying not to publish anything anywhere. When you are happy with the goal document, enter a name such as “Release researcher” in the “Name” field and press “Confirm goal.” To resume later, press “Continue later.” The draft remains.
✓The status becomes “Stopped,” and Work settings opens.
Review Work settings
Use the values below for this example. Each selection saves immediately.
| Field | Value for this example |
|---|---|
| AI to work with · Model · Reasoning effort | The same AI as the meeting (you can change it) |
| Work schedule | Daily at 09:00 |
| Work hours | 24 hours |
✓The settings chip shows something like “Daily 09:00 · 24 hours.”
Press “Start work”
The first shift starts immediately. The status changes to “Working · 1 minute,” then becomes “Resting · Next shift tomorrow 09:00” after the shift ends.
If work cannot start, “Why work cannot start” appears. Common reasons are that the working AI needs sign-in or the employee has no goal yet.
✓The list row shows “Working” or “Resting.”
View results in Journal and Memory
Open the latest shift’s card in the “Journal” tab. It shows a few lines of work recorded by the employee and the time taken. Expand “Actual actions · N entries” to see actions orke observed in the browser: URLs opened, buttons clicked, and text entered.
The “Memory” tab contains the status, plan, lessons, and journal files written by the employee. In its first shift, the employee makes a plan toward the goal, then starts working immediately.
✓The plan file describes what to do in the next shift.
Send an instruction
Enter what you want to say in the field at the bottom of the “Inbox” tab. Press “Send” for the next shift to read it. Check “Send and work now” before sending to start a shift immediately and read it. Instructions do not expire after one reading. Every shift keeps following them until you change them.
✓“Read · 09:00 shift” appears under the instruction, followed by the employee’s reply when the shift ends.
Handle requests when they arrive
When an employee reaches a step only a person can do, such as entering a password, verification, payment, or a CAPTCHA, it leaves a request and continues other work. For example, if a members-only forum requires email verification, it requests help with that site and checks other projects first. A “Release researcher · Left a request” notification appears on the PC, and a request bar appears at the top of the employee screen.
If the request includes a page URL, use “Open in the orke browser” in the request bar to open it. Complete the requested action yourself, then press “Done.” If you do not want to do it, press “Won’t do.” The employee does not make that request again and finds another way. To explain your thoughts, press “Give my opinion,” type, and send. The employee reads your message, decides whether the request was completed or declined, and replies in one line in the Inbox. In every case, it reads the result in its next shift and continues. Press “Work now” to have it continue immediately.
✓The Inbox keeps “Resolved · Done,” “Resolved · Decided not to,” or “Opinion · Your message,” and the request bar disappears.
Claude · Daily 09:00 · 24 hours · Next shift tomorrow 09:00
This example illustrates the workflow. Actual content depends on the employee’s goal and shifts.
Example 2: a promotion employee and requests
An employee that publishes posts and comments works differently. Under “What should it take on?”, enter something like “Find ways to spread the word about our service and promote it consistently.” Clearly define things not to do in the meeting. The goal document below includes a number and deadline because the person said “500 newsletter subscribers by January 10, 2027” during the meeting. If you do not specify them, the goal is set without a number or deadline.
---
due: 2027-01-10
metric: Newsletter subscribers = 500
---
## Goal
Introduce our service to developer communities and increase newsletter subscribers.
## Things not to do
- Do not pose as an ordinary user. Disclose that the account belongs to the service’s creator.
- Do not copy and paste the same post to multiple places.
## Facts to know
- The introduction page URL and three main features
## Metrics
Check subscriber counts in the newsletter management dashboard.- Places and methods — The employee looks beyond the places you suggest and independently finds more communities and methods suited to the goal. It decides where and how often to post.
- Sign-up and sign-in — If posting to community A requires an account, the employee uses the Google account signed in to the orke browser and presses “Continue with Google” to sign up and sign in. If sign-up needs SMS verification, it leaves a request saying “Signing up for community A requires SMS verification” and works elsewhere first. When you complete verification and press “Done,” it posts there in its next shift. Press “Won’t do” if you do not want to use that community; the employee finds another place.
- Checking metrics — When the employee checks the subscriber count, the metric card in the “Goal” tab shows “Current value / 500.”
Employees use your accounts signed in to the orke browser (Google, GitHub, etc.) to sign up for sites and independently publish posts, comments, and messages without approval. There is no daily posting limit. Many social networks and communities treat automated comments as spam and may block accounts. Recommendations that pose as ordinary users can also create problems as undisclosed advertising. Clearly state the rules to follow under “Things not to do” in the goal document during your meeting.
Meetings and the goal document
In the first meeting, the AI reads the assignment and checks public facts, such as your homepage, if needed. It fills out the entire goal document first, then asks just once: “Shall we start with this? Let me know if anything needs changing.” It does not ask for numbers and deadlines one by one; it includes them only if you specify them. It does not include metrics the employee cannot verify itself. If the assignment is broad, it keeps that breadth instead of narrowing it to particular places or channels. The goal document does not prescribe where, what, or how often to work. The employee decides how to get there during its shifts.
Only you can open meetings. Use “Open meeting” on the employee screen to reopen one at any time. You then choose the AI for the meeting. The employee’s working AI is selected initially. Choosing another meeting AI does not change the AI doing the work. If a previous meeting with the same AI exists, that conversation continues. In a reopened meeting, the AI briefly shows the current goal and asks only “What would you like to change?” It changes only what you mention and does not ask about everything else again. No new shifts start during a meeting, but a shift already running continues to the end. Pressing “Confirm goal” applies the new goal from the next shift.
| Goal document section | What to write |
|---|---|
| Goal | One or two sentences describing what to achieve. This is the goal statement shown in the app. |
| Things not to do | Rules the employee must follow. |
| Facts to know | Information the employee needs, such as product descriptions and URLs. Include source URLs for facts checked on the web. |
| Metrics | Include only when there is a numeric goal. State where the employee can check it. The number is in a metric: NAME = TARGET_NUMBER line at the top of the document. |
| Deadline | Include only when you specify one. Use due: 2027-01-10 at the top of the document. The app shows something like “90 days left.” |
Work settings
Press the settings chip at the top of the employee screen, such as “Claude · Opus · High · Every 2 hours · 24 hours,” to expand Work settings. You can change them while a shift runs; changes apply from the next shift.
| Field | Choices | New employee default |
|---|---|---|
| AI to work with · Model · Reasoning effort | Claude · ChatGPT (Codex) · NVIDIA · Antigravity. Reasoning effort is available only for Claude and ChatGPT. | The AI selected in New AI employee |
| Work schedule | Nonstop · Every N minutes · Every N hours · Daily at HH:mm | Every 2 hours |
| Work hours | 24 hours, or start and end times with “Set hours” (overnight ranges such as 22:00–06:00 are allowed) | 24 hours |
You do not set a daily count for posts, comments, or messages. There is no count limit. Put the rules the employee must follow in “Things not to do” in the goal document.
How a shift works
A shift is an AI run without an interactive screen. On each shift, the employee does the following.
- Reads the goal document, memory files, your instructions, and previous requests and their outcomes. It rereads instructions every shift until you change them, rather than following them for just one shift.
- Chooses the most useful work now and updates its plan. It also looks for new places and methods suited to the goal beyond those you suggested.
- Works in the orke browser. It signs up and signs in itself. When only a person can complete a step, it leaves a request and works on something else.
- Immediately records published posts, comments, and messages, and replies to new instructions in one line.
- Updates status, plan, lessons, and journal files, leaves a few lines about what it did, then finishes.
With Claude or Codex (ChatGPT) as the working AI, a single shift continues in the same conversation. If the AI finishes early but time remains within the shift’s limit (shift-timeout in employee.md, two hours by default) and it actually did work, orke tells it to continue with the next item in its plan. A shift therefore becomes an extended work period instead of a one-minute check. When continuing, orke also passes along instructions and resolved requests received in the meantime. NVIDIA and Antigravity run once per shift.
The work schedule determines the next shift time. If that falls outside work hours, it moves to the next work-hours opening. When shifts repeatedly finish quickly without doing anything, orke rests longer before the next shift (2 minutes → 4 minutes → 8 minutes → 16 minutes → up to 30 minutes). The same applies to a “Nonstop” schedule. During this time, the status shows something like “Nothing to do · Next shift 14:20.” Sending an instruction or handling a request clears the delay so work resumes soon. Failed shifts retry after 5 minutes → 30 minutes → 1 hour → 3 hours → 6 hours.
Employees whose work was enabled resume when orke starts again. Shifts missed while it was off are caught up only once. If orke stopped during a shift, that shift remains “Interrupted,” and the next shift checks the ledger first to see what was already done.
Sign-up, sign-in, and requests
Employees sign up and sign in to the sites they need themselves. They use buttons such as “Continue with Google” with accounts already signed in to the orke browser (Google, GitHub, etc.). Accounts created this way belong to you. Employees neither enter nor create passwords. If a site requires a password, verification, payment, or CAPTCHA, they leave a request for that site and keep working elsewhere.
- What they request — Only password entry; phone, SMS, email, and identity verification; payment; CAPTCHAs (tests that verify you are human); and tasks that truly require your account or action.
- What they do not request — Materials, information, numbers, permission, or direction. Employees find, create, and decide these themselves.
- One pending request — While a request is pending, the employee continues working without making another request. If all remaining tasks depend on you, it ends the shift and reads your answer in the next shift. It collects additional needs, then combines them into a single request after the pending one is handled.
- Three ways to respond — On PC or phone, choose “Done,” “Won’t do,” or “Give my opinion.” After “Won’t do,” the employee does not repeat that request and finds another way. “Give my opinion” sends your own words; the employee reads them, decides whether the request was completed or declined, and replies in one line.
Each shift, the employee reads past requests and their outcomes together.
Statuses and buttons
| Status | Meaning | Buttons shown |
|---|---|---|
| Draft | No goal yet. | Open meeting |
| Stopped | A goal exists, but work has not started or has been stopped. | Open meeting · Start work |
| Working | A shift is running now. Shows elapsed time, such as “12 minutes.” | Open meeting · Stop |
| Resting | Waiting for the next shift. Shows the reason, such as “Next shift 14:00,” “Outside work hours · From 09:00,” or “Nothing to do · Next shift 14:20.” | Open meeting · Work now · Stop |
| Waiting to retry | Waiting to retry a failed shift. | Open meeting · Work now · Stop |
| In a meeting | A meeting is open. No new shifts start. | (Meeting screen) |
| Setup error | Settings or the goal document could not be read. A reason is shown. | Open meeting |
The line below the employee’s name shows the reason and today’s recorded counts, such as “Next shift 14:00 · Today: 2 posts · 5 comments.” Types with no recorded actions are omitted.
The “⋯” menu contains Open folder, Rename, and Delete. Stop a running employee before deleting it. Deletion removes its goal, memory, shift history, and ledger.
Inbox — instructions and requests
The Inbox keeps your instructions, employee replies and requests, and resolution records in chronological order. Instructions can be up to 4,000 characters and remain valid until you change them. Every shift, the employee reads recent instructions together with past requests and their outcomes.
| Entry | Who | Meaning |
|---|---|---|
| Instruction | You | A message for the employee. Read from the next shift onward and followed until changed. |
| Read · 10:12 shift | orke | Which shift received that instruction |
| Reply | Employee | A one-line reply to an instruction |
| Request | Employee | Something only a person can do, such as a password, verification, payment, or CAPTCHA. Only one may be pending at a time. |
| Resolved | You | A record of “Done” or “Won’t do.” The entry shows “Done” or “Decided not to.” |
| Opinion | You | A message sent about a request with “Give my opinion.” A muted line below identifies the request. The employee reads it, decides, and replies in one line. |
Below the input field, a notice shows when the instruction will be read. During a shift, it says “Currently working; will read after the current work finishes.” While resting, it says “Will read at the next shift at 14:00.” To have it read immediately, check “Send and work now” before sending (the notice changes to “Sending starts a shift immediately to read it”). On a phone, press “Send and work now” beside “Send.”
Requests appear in the request bar at the top of the employee screen regardless of the selected tab, with “Done,” “Won’t do,” and “Give my opinion” buttons. “Give my opinion” opens a text field there. If a request includes a page URL, “Open in the orke browser” also appears. Notifications show only “Employee name · Left a request”; read the request text in orke. Handling a request does not wake the employee, so press “Work now” if you want it to continue immediately.
Journal, ledger, and memory
- Journal — Grouped by date, with recent shifts first. Each shift card shows its status (Working, Finished, Failed, Interrupted, or Stopped), the employee’s account of its work, time taken, and published counts (“41 min · 1 post · 3 comments”). URLs in the work description open in an orke web window.
- Ledger — “Actual actions · N entries” in the Journal. orke records browser actions as open, type, click, key, post, goal, step, and close. These are orke’s records rather than the employee’s claims, so you can compare its description with actual actions.
- Goal — Shows the goal statement, deadline, metric cards (current value / target and the shift that checked it), and remaining sections.
- Memory — Status, plan, lessons, and dated journal files. Use “Edit” to change them yourself; changes apply from the next shift. If the employee has changed the same file in the meantime, orke tells you instead of overwriting it.
Employee files are in ~/.orke/employees/<EMPLOYEE>/ (on Windows, %USERPROFILE%\.orke\employees\). Open them directly with “⋯ → Open folder.”
| File or folder | Contents |
|---|---|
employee.md | Work settings and assignment |
goal.md | The goal document confirmed in a meeting |
work/memory/ | Status, plan, lessons, and journal (the employee’s memory) |
work/ | The employee’s desk: outputs such as drafts and images it creates |
inbox.jsonl · ledger/ · shifts/ | Inbox, ledger, and shift history (written by orke) |
Things to keep in mind
- There are no count limits for posts, comments, or messages. Counts in the Journal and app (“Today: 2 posts · 5 comments”) are recorded by the employee immediately after publishing. Use the ledger to check whether it published anything without recording it.
- No separate usage-limit protection is enforced. A shift that hits an AI usage limit counts as failed and retries after a delay.
- Employees request help rather than entering passwords; completing phone, SMS, email, or identity verification; making payments; or solving CAPTCHAs. They sign up and sign in directly with accounts already signed in to the orke browser. The signed-in state of orke web windows is shared with employees and other agents. Sites an employee signs up for or signs in to also remain signed in within that browser profile.
- Notifications arrive only when you need to act: a request, work blocked by required sign-in, the working AI not being found, or repeated shift failures (the fifth consecutive failure).
Use employees from your phone or the CLI
Phone — Connect to your PC’s orke with the orke app for iPhone or Android to see the “AI employees” group in the menu. Select an employee to use “Done,” “Won’t do,” and “Give my opinion” on request cards; view the Journal and Inbox; send instructions; and use “Start work,” “Stop,” and “Work now.” “Send and work now” is available only when work is enabled. Use the PC for goals, memory, work settings, and meetings. There are no phone notifications, so open the app to check. See Remote and mobile for connection instructions.
CLI — You can also manage employees from a terminal.
orke employee list
orke employee instruct --slug EMPLOYEE_SLUG --text "Include an original source link with every summary" --wake true
orke employee resolve --slug EMPLOYEE_SLUG --id REQUEST_IDTo handle a request as “Won’t do,” add --decline true to resolve. To send an opinion, add --text "MY_OPINION". See AI employee commands for all commands and options.
When problems occur
| What you see | Checks and next action |
|---|---|
| Work will not start | Read “Why work cannot start.” If there is no goal, set one in a meeting. If sign-in is required, sign in to the working AI in Settings. |
| “Could not work because sign-in is required” notification | Sign in to the working AI (Claude Code, Codex, etc.) again, then press “Work now.” |
| Repeatedly waiting to retry | Check the reason in the failed shift’s Journal card. If it is a usage limit, work resumes when the limit clears. |
| “Nothing to do · Next shift 14:20” | Shifts have repeatedly finished quickly without work, so orke increased the rest period (up to 30 minutes). Check the Memory tab’s status file and Journal to see why the employee thinks there is nothing to do. Send an instruction if needed. Sending an instruction clears the delay. |
| Instruction has not been read | Check the notice below the instruction. If work has not started, start it so the instruction can be read. To have it read immediately, use “Send and work now.” |
| The same request returns | Check that you actually completed the requested action (for example, that the signed-in state remains in an orke web window), then press “Done.” If you do not want to do it, press “Won’t do.” The employee does not repeat that request and finds another way. |
| Setup error | Fix the file and line given in the reason. If you edited employee.md directly, check its format. |
Common points of confusion
When should I use an AI employee instead of a recurring task (package)?
Recurring tasks suit work whose sequence is clear and the same each time. An AI employee suits a clear goal where the next task changes and past results need to be remembered and built on.
Can an employee publish on its own?
Yes. It publishes without approval and has no daily count limit. To prevent publishing, state that in “Things not to do” in the goal document during a meeting. Published activity is recorded in the Journal and ledger.
Can an employee sign up and sign in?
Yes. It uses buttons such as “Continue with Google” with Google, GitHub, and other accounts signed in to the orke browser to sign up and sign in directly. It does not enter or create passwords. Sites requiring a password, verification, payment, or CAPTCHA become requests while it works elsewhere.
Will an employee ask to hold a meeting?
No. Only you open meetings with “Open meeting.” When you want to change the goal, open one. The AI shows the current goal and asks only “What would you like to change?”
Can multiple employees work at once?
Yes. Each employee has separate goals, memory, and work settings, and works independently without waiting for the others.
Does changing the working AI erase memory?
No. Memory is stored in files in the employee folder, so it carries over when you change the working AI.