When Anthropic's Fable 5 model was internationally suspended for nearly three weeks, many businesses overly reliant on it faced chaos. However, the 114-year-old insurance giant Liberty Mutual easily rerouted to other platforms thanks to a flexible "AI backbone" system built 18 months prior. Speaking at a recent VB Impact event, Brian Craig, Liberty Mutual's Senior Director of Architecture, emphasized the importance of avoiding vendor lock-in or fixed frameworks in a rapidly changing technological landscape.
Detailed Developments
Liberty Mutual's core architecture operates on the principle of separating an independent control plane from interchangeable execution layers. Thanks to this design, when the Fable 5 incident occurred, the company experienced no operational disruption, incurring only minimal time to re-route workloads to alternative models. This flexibility is further reinforced by a short-term contracting policy. Instead of signing traditional five-year enterprise agreements, Liberty Mutual proactively shifted to one-year contracts to easily evaluate and change vendors in line with the evolving pace of the AI market.
Technical Analysis & Technology
Liberty Mutual's AI backbone system comprises approximately 50 independent components, ranging from security, identity, orchestration, and rate limiting tools to agent behavior management policies. Beneath this backbone, they leverage AWS's Amazon Bedrock AgentCore as the agent runtime environment due to its multi-framework and multi-model support capabilities. Notably, instead of using a single agent to handle all tasks, which often leads to context overload, Liberty Mutual designed a "software factory" consisting of six sequentially collaborating agents:
* Epic agent: (receives requests) * Story agent: (breaks down work) * Planning agent: (technical planning) * Coding/Testing agent: (programming and testing) * Triage agent: (quality assessment) * Librarian agent: (knowledge context search)
Expert Opinion & Commentary
According to Brian Craig, role decomposition helps narrow the context window and allows for much more precise control over AI outputs. Initial deployments have shown the system completing three months' worth of work in just one week by eliminating most human handover waiting times. Nevertheless, Craig affirmed that Liberty Mutual does not operate a fully automated process but maintains an iterative cadence with human oversight. All results from the software factory must pass engineer approval before being moved into production.
Impact & Future
Liberty Mutual's multi-cloud and model-agnostic architectural approach offers a valuable lesson for Vietnamese businesses developing AI solutions. Integrating deep observability into the backbone ensures tight control over the data that agents access, preventing AI from operating outside of established bounds. In the finance and insurance industry, which demands absolute security, maintaining human review combined with a flexible technological platform is key to both fostering innovation and ensuring robust risk management.