Resilience Checklist

AI is software, and the lessons learned from decades of software development and deployment remain relevant for AI agents. Check that these lessons are applied to your proposed or existing agent...

Does your agent...
get frequent evaluations against a broad range of edge cases?
gracefully handle incomplete, ambiguous, or contradictory inputs?
detect and recover from retrieved data that is corrupt?
still respond sensibly when its APIs or databases are unavailable?
have failover mechanisms for its models, data sources and APIs?
degrade gracefully under high load or constrained resources?
detect and handle adversarial attacks?
monitor its own performance and autonomously escalate any problems?
retry failed tasks using alternative agent pathways?
log its failures in detail to improve future resilience?
get comprehensive evals after updates to models, data, or software?
function correctly on all intended platforms and channels?
redirect to a human when all else fails?

Resilience Risk: 100

Unless you've addressed most of the items in the above checklist, the risk that your agent will fail is high. Check the boxes to answer the above questions and get an assessment of your real risk.

Our understanding of AI agent resilience is evolving quickly. Be sure to check this list regularly for new knowledge.