AWS re:Invent 2025 - Build useful, reliable agents with Amazon Nova (AIM372)

Summary of AWS re:Invent 2025 - Build useful, reliable agents with Amazon Nova (AIM372)

Transition to Agentic Systems

• Customers are moving from models that generate insights to models that can take action, known as "agents"
• Agents are task-oriented, purpose-built AI systems that can complete work in enterprise systems, bridging the gap between intelligence and execution
• As this transition continues, we're seeing the rise of multi-agent systems that can take on more complex tasks, coordinate, and work towards common goals

Key Layers for Building Agents

1. Agentic Primitives: The building blocks that allow agents to interact with the world, including tool orchestration, memory management, and observability
   • Amazon offers Bedrock Agent Core to enable secure, scalable agent deployment
2. Models: Not all models are equally capable for agentic workflows; models need to be able to reason through complex tasks, break down plans, and reliably call tools
   • Amazon's Nova 2 models are designed specifically for agent-based workflows
3. Multi-Agent Frameworks: Complex real-world use cases often require multiple specialized agents working together
   • Frameworks like Strands provide structured approaches for multi-agent interactions and orchestration

Capabilities of Nova 2 Models for Agents

1. Tool Calling: Nova 2 models can reliably select the right tool, pass the correct parameters, and understand the tool's output to plan the next steps
   • Includes built-in tools like Code Interpreter and Web Grounding
2. Multi-Step Reasoning: Nova 2 models can break down complex tasks, understand plans, and adapt when things go wrong
   • Customers can choose the level of reasoning (no, low, medium, or high) based on their use case needs
3. Extended Context: Nova 2 models can process large amounts of data (up to 1 million tokens) to understand relevant information for real-world workflows

Reliability and Robustness of Nova Act

• Nova Act combines the model, orchestration, memory, and compute components into an end-to-end solution focused on high reliability
• Key capabilities:
  • Understanding and navigating complex UIs, not just APIs and deterministic flows
  • Learning cause-and-effect through reinforcement learning in simulated environments
  • Achieving over 90% reliability in real-world enterprise workflows
• Enables use cases like:
  1. Form filling and data entry
  2. Searching and extracting information from secured enterprise systems
  3. Automated booking and checkout flows
  4. Scalable UI-based QA testing

Multi-Agent Workflows with Strands

• Multi-agent frameworks like Strands enable specialized agents to work together on complex tasks
• Benefits of multi-agent systems:
  1. Specialization: Applying the right model/agent for each subtask
  2. Scalability and Modularity: Easily adding or upgrading agents without rewriting the entire workflow
  3. Latency and Efficiency: Parallel execution and using the most appropriate model for each step
• Example: Sumo Logic's security operations agents using Nova reduced resolution time by 75%

Key Takeaways

• The future is "agentic" - models that can take action and complete real-world tasks
• Nova 2 models and Nova Act provide powerful capabilities for building reliable, high-performing agents
• Multi-agent frameworks like Strands enable specialized, modular, and efficient agent-based workflows
• Customers are already seeing significant benefits in areas like security operations, QA, and customer service automation