AWS re:Invent 2025 - Powering AI and agentic platforms with Amazon Bedrock (AIM3305)

Powering AI and Agentic Platforms with Amazon Bedrock

Challenges in Going from Prototype to Production

• Companies cancel an average of 46% of their AI projects before reaching production
• Key reasons for project cancellation:
  • Platforms are too static and can't keep up with the pace of change
  • Concerns about non-deterministic model outputs and how to manage them
  • Lack of visibility into application issues
  • Difficulty integrating AI into traditional software development lifecycles

Pillars of an AI Platform

The key pillars of a robust AI platform include:

1. Models: Access to a catalog of diverse models, including frontier models from providers like Anthropic, Nvidia, and OpenAI.
2. Deployment and Orchestration: Secure and scalable deployment and configuration management of models and agents.
3. Data Foundation: Integrating relevant data sources, knowledge bases, and vector databases to provide context.
4. Security and Governance: Implementing safeguards, content filters, and policies to ensure safe and responsible AI.
5. Agentic Capabilities: Building and deploying intelligent agents that can interact with models and data sources.

Enabling Model Switching

• Importance of being able to easily switch models as new ones are released or for different use cases
• Challenges include varying model interfaces, lack of a plug-and-play approach, and difficulty evaluating model performance
• Amazon Bedrock provides tools to enable model switching:
  • Converse API: Unified interface for calling any Bedrock model with consistent input/output formats
  • Strands Agent SDK: Allows agents to be model-agnostic, enabling easy model swapping
  • Bedrock Guardrails: Configurable safeguards and content filters applied to model inputs and outputs

Evaluating Model Performance

• Importance of being able to evaluate model performance for your specific use case
• Challenges include finding the right metrics and algorithms, sourcing relevant data sets, and setting up evaluation infrastructure
• Amazon Bedrock offers tools to simplify model evaluation:
  • LLM as a Judge: Using a large language model to evaluate other models or agents
  • Bring Your Own Inference: Ability to evaluate models and applications hosted outside of Bedrock
  • Agent Core Evaluations: Continuous real-time scoring of agent performance with built-in and custom evaluators

Observing AI Application Behavior

• Importance of observability to assess system health, root cause issues, and track performance degradation
• Challenges include fragmented tracing, scaling evaluations, and lack of visibility into agent performance
• Amazon Bedrock integrates with observability tools:
  • Amazon CloudWatch: Out-of-the-box insights into application performance, health, and accuracy, with end-to-end prompt tracing
  • Agent Core Observability: Comprehensive view of agent behaviors and operations, with real-time dashboards and integration with third-party tools

Verscell's Approach with Amazon Bedrock

• Verscell, a provider of open-source frameworks and self-driving infrastructure, uses Amazon Bedrock as the foundation for their AI and agentic workloads
• Key components of Verscell's AI platform:
  • Verscell AI SDK: Abstraction layer that enables writing application logic once and plugging in different models
  • AI Gateway: Provides detailed visibility into agent interactions with models across different providers
  • Workflow Development Kit: Allows building durable, long-running AI workflows that can handle asynchronous tasks

Key Takeaways

• Amazon Bedrock provides a comprehensive set of tools and services to help organizations build and manage robust AI platforms
• The ability to easily switch models, evaluate performance, and observe application behavior are critical for successful AI deployments
• Verscell's approach demonstrates how Bedrock can be leveraged as a foundational platform for building intelligent, agentic applications