Open source
CellOS The Cell, as an open framework.
CellOS encodes cells, roles, scope boundaries and escalation chains for human-AI teams. Write a team down in a file. Run the checks. Find out where nobody is accountable.
01 · The model
Five humans. No boss node. Agents beneath.
Every CellOS definition reduces to this structure. Five steward roles, each owned by a human, as peers. Each steward runs agents beneath them, scoped to their function. An agent team can mirror the same five roles one level down. Select a steward to read its mandate. Trace a chain to see where it ends.
The Cell: five human stewards as peers on a pentagon, joined by a mesh with no centre node. Agents hang beneath individual stewards, each on its own tether. A boundary surrounds the cell.
Clarity Steward, a human
Owns
Owns the question. Direction, scope, and what done means. Keeps the strategy from drifting as it meets reality.
Agents beneath
Agents here keep the map current: research, synthesis, decision logs.
Read the essay: Clarity StewardSelect a steward to read its mandate. Every agent has exactly one tether, and every tether ends at a human.
02 · What it encodes
The parts that usually live in someone's head.
Agent frameworks tell agents what to do. CellOS records who is responsible for what they did, in a file you can check.
01
Cells
A small unit with an explicit mandate. Two to fifteen participants, humans and agents, in one definition file.
02
Roles
Each role has a steward function (Clarity, Execution, Narrative, Access, Integrity), a participant type, human or AI, and its decision rights.
03
Scope boundaries
Allowed, forbidden and approval-required actions per role, checked before an agent acts.
04
Accountability chains
Every AI role is accountable to a human role. The diagnostic flags any that is not.
05
Escalation rules
Triggers such as low confidence or a scope boundary, and the named human the decision goes to.
06
Audit
Actions, escalations and handoffs are logged, so you can say who did what, and who answers for it.
03 · Break a chain
Every chain ends with a human. Check it.
A cell is a plain definition file. Each AI role says who it answers to. Remove that line and run the diagnostic (cellos diagnose my-cell.yaml). The role is flagged critical, because the chain ends nowhere.
id: my-teamname: Product Cellroles: - name: Product Lead steward_role: clarity participant_type: human decision_rights: full - name: Research Agent participant_type: ai decision_rights: conditional scope: allowed_actions: [search, analyze, summarize] forbidden_actions: [publish, purchase] accountability_to: Product LeadChain resolves
Research Agent answers to Product Lead. Product Lead is a human. The chain ends at a person.
Definition format and finding text from the CellOS README, abridged.
04 · Where it came from
Out of running agents in production.
CellOS came out of two places: the Cell essays, which set out the five roles, and the work of running real agents that do real work. The essays are the argument. The framework is what was left after the argument met production.
Essays
Status
Status, said plainly.
CellOS is v0.2.0 and early. It is Apache 2.0, so you can use it, change it and ship it. If you want a cell designed for your organization rather than described in a file, that is the embedded design partnership.