Liberty Mutual's model-agnostic AI backbone weathered the Fable 5 outage

When Anthropic's Fable 5 was pulled from international use for nearly three weeks, some over-reliant businesses were left scrambling. Liberty Mutual easily pivoted to other platforms because they built an AI backbone exactly for this scenario 18 months earlier.

What happened

Liberty Mutual built an independent AI control plane (the

FAQs

What is Liberty Mutual's AI backbone and how does it work?

It is an independent AI control plane that manages multiple interchangeable AI models and platforms. The AI backbone uses AWS Bedrock AgentCore for agent execution and a network of roughly 50 interoperable components across security, identity, orchestration, policy, and tool restriction. Observability and governance are baked in through Microsoft Entra ID for identity and access, MLflow for evaluating model outputs, and a context repository that guides agent decisions.

How does Liberty Mutual switch between AI models and platforms without lock-in?

The backbone is model-agnostic and designed for easy model and platform swaps through its interoperable components. Liberty Mutual also adopted one-year contract terms instead of the industry-standard five-year deals, enabling rapid evaluation and vendor changes as the market evolves.

What benefits did Liberty Mutual gain from a model-agnostic, multi-vendor architecture?

The architecture delivered faster onboarding and production ramp time, from months to roughly a week. It reduced dependency on any single vendor, allowing the team to route workloads based on price versus performance rather than vendor inertia. It also improved observability and governance over data access and tool usage, critical for a regulated insurer.

What roles do the six agents play in Liberty Mutual's AI factory?

The six specialized agents work in tandem: an epic agent consumes high-level requirements, a story agent breaks work into narrow slices, a planning agent defines the execution plan, a coding/testing agent handles coding and review, a triage agent reviews quality across all agents, and a librarian agent manages the knowledge context for the others. Separating them shrinks context windows and improves output quality.