How it enables collaboration
The unit of work is the contract that crosses every handoff: definition of done, budget, identity, state, and record.
Collaboration is a handoff problem. The unit of work solves it the way good interfaces always do: by standardizing what crosses the boundary.
Collaboration, whether human-to-agent or agent-to-agent, is a handoff problem. The question is always: what exactly travels across the boundary between two workers?
Without a shared unit, what travels is conversation. Context degrades at every hop, intermediate state lives in nobody's hands, and accountability blurs as the task passes through many hands. As the Phase 2 thesis puts it: a fleet without an owner is not an organization, it is noise.
Figure 2: Without a shared unit, collaboration is prompts back and forth and context is lost at every hop. With units of work, each handoff passes a verifiable state plus the contract: definition of done, budget, identity, and record.
The unit of work converts the handoff problem into an interface problem.
A common contract. Every participant, whether human, agent, or team of agents, receives the same package: an end state, a budget, the current state of the artifacts, and the record so far. Nobody needs the full history of the conversation to pick up the work, because the unit carries its own context.
Identity across hops. Because the unit, not the chat session, is the persistent object, identity and attribution travel with it. When work crosses from Agent A to Agent B, the record shows who did what at every step, so the outcome stays attributable even when no single worker saw it end to end.
Parallelism without chaos. Decomposing a goal into units of work is what allows a fleet to attack it in parallel: units are independently owned, independently budgeted, and independently verifiable, so they can be executed concurrently and re-integrated by checking states rather than by reconciling narratives.
Humans and agents become interchangeable at the boundary. A unit of work is exactly the shape in which a manager already delegates to a person: scope, budget, deadline, check the result. Keeping the same shape for agents means a mixed human–agent org needs one management protocol, not two. People do not leave the loop; they hold the units that need judgment, and route the rest.