Degradation Ladder

Define an agent system's fallback modes in advance (full service, reduced tools or cheaper model, read-only, paused with a holding message, stopped) and classify workloads by droppability while calm, so retreat during an incident is chosen rather than a collapse.

Evidence: weak

An agent system that is either fully up or fully down gives its operators, and any agent supervising it, two choices, and both are poor during a partial failure. Keeping full service running risks compounding damage; stopping everything sacrifices work that could safely continue. Working out where to land in between, in the middle of an incident, is slow and error-prone.

The pattern comes by analogy from resilience engineering. When one line of defence is breached, a resilient system retreats to the next, lowering its ambitions and sacrificing less-critical goals; the lines are identified beforehand so that retreat is a chosen runbook step rather than an unplanned collapse. Transferred to agent systems, the lines become a ladder: full service, then a reduced tool set or cheaper model, then read-only or answer-only, then paused with a holding message, then stopped. Work on self-care under limited capacity adds the timing argument: demands should be sorted green, yellow and red while headroom exists, because judging what to drop under low capacity is itself expensive.

One software-native source supports the middle rungs. Reliable Machine Learning describes switching to a simpler, bounded, non-optimal behaviour when neither rollback nor roll-forward is available, and warns that the fallback path must be built and exercised before it is needed. The ladder's shape and the traffic-light classification remain analogical, so the evidence is weak.

In practice each rung needs a concrete, pre-built configuration, and each workload (channels, scheduled jobs, individual agents) needs a droppability class decided ahead of time. Stepping down a rung then becomes a pre-approved action inside the agent's autonomy envelope, usually beginning with stopping new intake, while a hardline command floor keeps state and secrets protected on every rung. Rungs should be rehearsed, as fault-injection practice does for repairs. Repeated descents should feed supervision-tree escalation rather than oscillate silently.

Sources

  • Pariès, ch. 2, on defence in depth and tactical retreat
  • Neff, ch. 3, on precommitted demand-shedding
  • Chen et al., ch. 9, on fallback to a safe default