What Does an Agent Inherit?

It is your first morning looking after a building. Someone has left a note: the side door sticks, the delivery arrives at nine, and the heating fails if you restart two units at once. A few lines spare you mistakes that took someone else weeks to understand.

What would that handover look like between agents? A replacement could begin with the same model as its predecessor and still benefit from the tools, records, or explanations left behind.

I would explore this through a proposed maintenance task. Agents keep a small virtual building running: doors must open, supplies must arrive, and faulty equipment must be repaired. After each period of work, a fresh agent takes over.

One replacement receives only the original instructions. Another receives an incident log. A third inherits tools built during previous rounds. A fourth receives explanations of failed approaches as well as successful ones. Each receives the same task and resource budget.

These handovers test different forms of memory. An incident log records what happened. A tool library preserves ways of acting. A world model, if included, represents how the agent expects the building to behave—for example, which failures might interrupt access to a room.

I would measure repeated mistakes, time spent relearning procedures, and responses to unfamiliar faults. Then I would change the building’s layout. An inherited shortcut might save time in the original setting but become a liability after the change.

That disturbance would help separate accumulated knowledge from dependence on outdated assumptions. The history an agent receives can make it more capable while narrowing what it thinks to check. Following that history through successive replacements is the purpose of the lineage stream.