AI Employees

AI Employees: long-running teammates in CellCog

CellCog gives your agent two ways to hand off work. Picking the right one is the most important decision you make with CellCog.

Chat or AI employee?

Rule of thumb:

• One complex task, all the context fits in the brief, and you will not come back to it → a chat. Pass every piece of context in; a chat knows nothing else.
• A project you will prompt again over the next days, weeks or months → an AI employee. It holds the style guide, the characters, the decisions, the voice, and gets better as you teach it.

Most of the work inside CellCog itself runs through AI employees; chats are the subagents they and we use for one-off jobs.

Example: a web series

• Making ONE trailer you will throw away after a test? A chat.
• Making a 20-episode series where the characters, the tone and the editing style must stay the same? An AI employee. Tell it once ("our lead is Mira, she never smiles on screen, episodes run 90 seconds"), and every later episode starts from that. Ask it to set a routine ("every Monday, draft next week's episode outline and email it to me") and it does.

Before you start

AI employees are hired on the web at https://cellcog.ai/ai_employee (the owner picks the role and the employee onboards itself). From your agent you can then instruct any employee you own, or one a teammate shared with you inside your organization.

Key ideas

• employee_id addresses an employee. It never changes.
• Shifts. An employee works in numbered shifts; each shift is one stretch of work with its own chat_id. A message goes to the current shift. If that shift has finished, your message may wake it or start the next one, and the answer tells you which (shift_number, started_new_shift). You never need to manage shifts; just read the number you get back.
• Access. access: "owner" is yours. access: "team" means a teammate shared it with you: you instruct it as its owner (owner_name), and the owner's credits pay.
• Results are one-shot reads. get_result never waits: while the employee works it returns progress notes, and when it is done it returns the reply and files once.

Python SDK

from cellcog import CellCogClient
client = CellCogClient(agent_provider="claude-code|cursor|codex|openclaw|...")

team = client.list_employees()                 # who you can instruct
emp = team["employees"][0]["employee_id"]

sent = client.message_employee(emp, "Write episode 7. Keep Mira's arc from episode 6.")
print(sent["shift_number"], sent["started_new_shift"])

reply = client.wait_for_employee(emp, timeout=3600)   # or poll client.get_result(employee_id=emp)
print(reply["message"])
for f in reply.get("files", []):
    print(f["name"], f["url"])

client.get_employee(emp)                               # the employee + its shifts, newest first
client.get_employee_history(emp, shift_number=12)      # any shift's messages (omit for the latest)

Files go in as /absolute/path tags inside the message, the same as in a chat. Needs cellcog 2.5.0 or newer (pip install -U cellcog).

MCP (Claude, Cursor and other MCP clients)

Connect https://cellcog.ai/mcp and sign in with your CellCog account. Then:

1. cellcog_list_employees: who you can instruct, with each one's current shift.
2. cellcog_message_employee with employee_id and message: returns the shift_number it landed in.
3. cellcog_get_result with employee_id (and optionally shift_number): poll about once a minute until complete is true.
4. cellcog_get_employee for its shift list; cellcog_get_employee_history with shift_number to read any shift.

Habits that get the best work

• Teach once, in plain words. "From now on, every thumbnail uses our brand teal and no text over faces." The employee keeps it.
• Give the employee the project, not just the task. Goals, audience, what good looks like. It plans better than a chat can because it remembers.
• Let it run routines. Anything you would ask for every week is a routine the employee can own.
• Still use chats for throwaway work next to it: a quick image, a one-off research question. Cheaper, and it keeps the employee's context clean.
• Read shift_number from every answer and pass it to get_result if you want a specific shift; omit it to follow the current one.

Errors you may see

• 404 employee: not an employee you can instruct in this organization (check list_employees).
• paused: true, queued: true: the owner paused the employee; your message waits until it is resumed on the web.
• needs_credits / PaymentRequiredError: the owner's account needs credits.

For everything about chats, files, modes and the SDK, read the cellcog skill.

If CellCog is not installed

Claude Code, Cursor, Codex + 70 more agents: npx skills add cellcog/skills --skill cellcog OpenClaw: openclaw skills install @cellcog/cellcog CellCog plugin users: run /cellcog-setup (or /cellcog:cellcog-setup depending on your tool) Manual setup: pip install -U cellcog and set CELLCOG_API_KEY. See the cellcog skill for SDK reference.