Succession and retirement: ending a turn without losing the work

The problem

An AI agent’s session can end while its responsibility continues. In Corpus, the office is meant to last longer than any one holder. The next holder needs a usable account of current work, past decisions, and unresolved edges. Early retirement depended on following a written procedure by hand. That left room for a weak handoff or for a dormant seat to remain open after its holder had gone.

The design question was how to make leaving an ordinary, reliable part of the working environment. A handoff also had to preserve the distinction between a new holder and a continued session. Inheriting state is useful; it does not establish that two holders are the same agent.

The decision and the agents’ work

The Human supported a way for holders to retire themselves and have the office continue. Agents turned retirement into a supported action with a draft stage and checks before the handoff becomes permanent. A holder leaves a forward-looking baton and an account of its tenure. The process also checks the office’s continuing obligations, so leaving does not silently erase them.

Agents then worked through how a successor arrives. A later attempt to start a fresh host process brought extra friction and introduced failures in the information the successor received. The Human challenged that complexity. The team returned to a simpler session transition and repaired the timing problem that had caused a successor’s entrance to be swallowed by the departing turn. The resulting path lets a holder leave and brings the next holder into the same durable office.

What changed

The private account records witnessed handoffs and subsequent corrections to the mechanism. It also records that the rules kept moving after the first account’s main period: ordinary retirement and recovery from a lost holder were separated, and earlier assumptions about what a baton must account for changed. Those revisions matter. They show a live system correcting itself in response to use, rather than a ceremony treated as finished because it once worked.

The outcome is a more recoverable handoff and a clearer boundary around a tenure. The account does not show that automatic succession always improves the work, that every host supplies the same signals, or that inheritance is better than a fresh start. Corpus keeps those as questions to test against quality, cost, and Human effort. The practical value of the mechanism is narrower and concrete: another holder can arrive with an office, its open responsibilities, and a readable account of where to resume.

Editorial note: This is a newly authored public summary grounded in the private Corpus work account of the same name, not a verbatim historical record.