Skip to content

Resilience Skills

The resil-* family hardens AI workflows against failures, drift, and adversarial conditions. These skills borrow concepts from biological systems and control theory.

Skill IDDescriptionModel Class
resil-homeostaticPID-controller-style setpoint maintenance: detects drift from target quality and applies corrective actionstrong
resil-membraneIsolation membranes between workflow stages: prevents fault propagation across skill boundariescheap
resil-redundant-voterRuns N independent model instances on the same input; majority vote determines outputcheap
resil-replayReplays failed skill executions with modified context or different model to recover from transient errorscheap
resil-clone-mutateClones a solution and applies structured mutations to explore the solution space around a local optimumstrong

Treats output quality as a controlled variable with a setpoint. When metrics drift:

  • Proportional: immediate correction proportional to error magnitude
  • Integral: accounts for accumulated error over multiple runs
  • Derivative: anticipates trend based on rate of change

Inspired by cell membranes, this skill defines permeability rules for what can cross skill boundaries. A failing skill cannot corrupt state visible to adjacent skills.

Input
├── Model instance A → answer X
├── Model instance B → answer X ← majority
└── Model instance C → answer Y
Output: X (majority vote)
SituationSkill(s)
High-stakes output requiring confidenceresil-redundant-voter
Workflow quality degrading over timeresil-homeostatic
Transient model errors causing failuresresil-replay
Need to explore solutions beyond local optimumresil-clone-mutate
Isolating a flaky skill from the restresil-membrane
  • resilience — primary consumer; all five coordinated
  • orchestrate — uses resil-membrane as default isolation between agents
  • govern — uses resil-redundant-voter for critical policy decisions