· SYSTEMS
The Issue Is the Primitive. Running companies of humans and AI agents on Linear’s model
01
The real bottleneck
In my experience, long-horizon agents rarely fail at the step. They fail at the thread: what was decided, by whom, what is blocked, what done means. That state must live outside the model.
02
Linear’s primitives
Linear’s handful of concepts read as agent architecture. An issue is the unit of work: a bounded task with its own context and history. A team’s workflow, the ordered statuses an issue moves through, is a state machine. Triage, where new work is reviewed before acceptance, is an admission gate. Projects, milestones and initiatives (the outcome, its stages, the why) form a goal hierarchy. Cycles give a team a heartbeat. Views are queries over shared state.
03
Assigned versus delegated
What convinced me is how Linear added agents. An issue can be assigned only to a person but delegated to an agent. As Linear’s guidelines put it, an agent cannot be held accountable.
It echoes a 1979 IBM training slide: a computer can never be held accountable, so it must never make a management decision. Delegation puts that rule in the data model, not a policy document.
04
Memory outside the model
OSWorld 2.0 found that agents forget early lessons that live only in compressed reasoning. An issue is memory outside the model: written down, versioned, shared. Any agent can pick it up cold, and any human can audit it.
Conway’s law says organisations ship systems that mirror their communication structure. If agents communicate through issues, the company mirrors its tracker, and its structure becomes legible.
05
The playbook
My working rule: every unit of agent work is an issue, every agent has a human owner, every team owns its workflow, and nothing enters without triage. Scale by adding agents under that grid, not by giving one agent a bigger context window.
The tracker stops being a to-do list. It becomes the company.