AWS re:Invent 2025 - What Database Would Your AI Agents Choose - Escape the Frankenstack (DAT203)

Summary of AWS re:Invent 2025 Presentation: "What Database Would Your AI Agents Choose - Escape the Frankenstack"

The Frankenstack Problem

• Many organizations end up with complex "Frankenstack" architectures, with multiple disparate data stores, caches, streaming platforms, and custom ETL pipelines.
• This complexity happens gradually through a series of reasonable individual decisions, not a single bad choice.
• Frankenstack architectures lead to issues like:
  • Unpredictable latency
  • Cache inconsistency
  • Schema drift
  • Surprising costs
  • Difficult debugging and maintenance

The Impact of AI Agents

• AI agents amplify the problems of Frankenstack architectures, as they require real-time access to structured, unstructured, and vector data across multiple systems.
• Even a single slow or unreliable component can cause the entire system to collapse.
• Maintaining the web of interdependencies in a Frankenstack becomes a significant burden, requiring experts in each data store and constant firefighting.

Rethinking the Architecture for AI Agents

• Instead of building for dashboards and batch processing, organizations should design their architecture around the needs of interactive, stateful AI agents.
• Key requirements for AI agents include:
  • One unified data store with minimal data movement
  • Faster development patterns with fewer dependencies
  • Fewer background processes and monitoring tasks

Benefits of a Unified Data Store

• Measurable improvements in development velocity, reliability, and cost:
  • 3x fewer lines of code
  • Fewer database connections and failure points
  • Simpler debugging, CI/CD, and on-call rotations
  • Easier scaling and reduced operational overhead

Business Impact

• Faster feature delivery, as product teams no longer have to coordinate across multiple engineering teams and data stores.
• Increased reliability and predictability, as the system becomes a well-lit, maintainable platform instead of a haunted Frankenstack.
• Reduced operational costs and engineering burden, as the need for specialized experts and constant firefighting is eliminated.

Real-World Example

• The presenter shared a demo showcasing the differences between a Frankenstack architecture and a unified data store approach.
• Key metrics showed 3x fewer lines of code, fewer database connections, and fewer failure points in the unified data store solution.

Conclusion

• Frankenstack architectures are a common problem, but they happen gradually through reasonable individual decisions, not a single bad choice.
• AI agents amplify the issues of Frankenstack architectures, requiring a rethinking of the data architecture.
• By adopting a unified data store approach designed around the needs of AI agents, organizations can achieve significant improvements in development velocity, reliability, and operational costs.