← All agent and harness briefs
Agent runtimeautonomousopsresearchgtmdocumentedunattended capable

Operational brief

CellCog

Managed AI employees with persistent inboxes and boards, scheduled shifts, and user-configured signoff rules.

No product visual on file

Choose it when

You want a hosted agent to hold a standing role, retain work across shifts, and return decisions for approval rather than wait for the next prompt.

Skip it when

You need a local runtime, a conventional prompt-by-prompt assistant, or independently validated operational performance before deployment.

01 / Classification

Autonomy evidence

This is our classification, not a vendor badge. The record says what the product normally does. The gaps stay visible.

How we classify it
Unattended scheduler/heartbeat with write-tools; failure has a real owner.
Runs unattended
Yes — that capability is documented in this brief.
Trigger
The employee launch documents owner-set schedules and wakes from inbound email, a due task or a user message. These are employee capabilities, not a claim about every SDK call.
State
Employee email threads, task boards and memory persist across working sessions; the employee writes a handover for the next session.
Tools
The launch describes sending/receiving real email and producing code, dashboards, spreadsheets, PDFs and media. Connected-system actions depend on granted permissions and separately configured signoff.
Human control
CellCog describes one approval level across computer, browser and tools, not unified technical access grants. Above the employee's level, a quoted active standing approval can allow immediate execution. Separately, an authorized person at the employee's desk, in chat and having written within about an hour, can request or say yes to dangerous work; it's recorded as approved in chat and runs without a second card. The owner, shared Team members and enabled managers have specified approval rights. If the person cannot approve, the step doesn't run and the owner gets a task. Otherwise pending approval can wait for an active user for 15 minutes before failing; when away it fails immediately with guidance and an owner task. Standing grants may be scoped and time-limited: the platform documents existence, activity and expiry checks, while observing wording and limits is an expected employee contract, not demonstrated semantic enforcement. Outgoing employee email has a separate dangerous-operation policy and no approval button. Risk classification and enforcement are vendor-described, not independently tested. A human owner still monitors waiting work and failures; recovery and concurrency aren't guaranteed.

02 / Operations

Control reality

Useful agents eventually touch something real. This is the operational picture we have, not a promise hidden in a sales page.

Where it runs
cloud
Data boundary
Not recorded yet.
Identity & permissions
Not recorded yet.
What it can write
role task boards · email and connected work under granted permissions · handover notes and deliverable files
Surfaces
Hosted chat · employee email inbox · task board and channels · scheduled shifts
Export / lock-in
Not recorded yet.
Operational limits
Vendor documentation, not a hired-employee trial. A shared approval level across surfaces doesn't unify technical access grants. An active standing approval can allow above-level employee work; an authorized recent chat request or yes can also approve a dangerous step without a second card. Owner, shared Team members and enabled managers have specified rights; if the person cannot approve, the step doesn't run and the owner gets a task. Otherwise pending approval for an active user fails after 15 minutes without a decision; away behavior fails immediately with guidance. The documented standing-grant checks cover existence, activity and expiry; wording and limits are an expected employee contract, not demonstrated semantic enforcement. Outgoing employee email has a separate dangerous-operation policy and no approval button. Risk classification, defaults and actual policy enforcement remain unverified. This isn't a guarantee that every dangerous action is blocked. The permissions/shifts guides recommend recovery, concurrency, retry and audit contracts; those aren't proof of enforced defaults or a recovery SLA. Hiring/trial can be free; work is usage-based. Employee SDK support isn't established by these sources.
Watch for
Name a human role owner to review pending approvals, feedback and failures. Above-level employee actions may run on an active standing grant; an authorized recent chat request or yes can also approve a dangerous step without a second card. Owner, shared Team members and enabled managers have specified rights; if the person cannot approve, the step doesn't run and the owner gets a task. Otherwise an active user's pending approval fails after 15 minutes without a decision; away behavior fails immediately with guidance. Check standing-grant scope and expiry: documented checks cover existence, activity and expiry, not demonstrated semantic compliance with wording or limits. Outgoing employee email has a separate dangerous-operation policy and no approval button. The shared level doesn't unify technical access grants; check connected-system permissions. Classification and actual enforcement haven't been independently tested. Define and verify retry, overlap, spend and recovery rules before live writes; the vendor's design templates aren't an implementation guarantee.

03 / Receipts

Evidence ledger

Here is what this brief is built on. “Documented” means we found a source. “Tested” means this desk actually ran it. Those are different things.

Evidence status
documented
Source set
6 primary sources
Evidence review
2026-11-17

Corrections

  • 2026-10-14 · Replaced universal dangerous-action approval wording with user-configured signoff. Added the employee schedule/state/write/owner bundle, qualified free hire/trial versus usage-based work, and removed the unsupported employee Python SDK surface. Retained June 21 launch: the September 29 article update isn't a newer product release.
  • 2026-11-17 · Corrected exclusive signoff wording: CellCog describes a shared approval level, not unified technical access grants. Above-level employee work can use an active standing approval; authorized recent chat instructions or yes can also approve a dangerous step without a second card. Owner, shared Team members and enabled managers have specified rights; if the person cannot approve, the step doesn't run and the owner gets a task. Otherwise active-user pending approval has a 15-minute no-decision timeout and away behavior fails immediately with guidance. Standing grants can have scope and expiry; documented existence, activity and expiry checks are distinct from wording and limits the employee is expected to honor. Outgoing employee email has a separate dangerous-operation policy and no approval button. Risk classification, defaults and actual enforcement remain unverified; no desk test or security guarantee.

04 / Change log

Related coverage

The brief tells you what the product is. Payloads track the changes that may alter the decision.

No linked change coverage on file yet.