Building Autonomous, Self-Improving AI Agents with Strands

Overview

This presentation from AWS re:Invent 2025 showcases the capabilities of Strands, a framework for building autonomous, self-improving AI agents. The speakers, Aaron and Shagatai, demonstrate how these agents can dynamically create their own tools, update their system prompts, learn from interactions, and even orchestrate the creation of sub-agents to tackle complex tasks.

Key Features of Strands Agents

Self-Extending Agents

• Strands agents can create their own tools during runtime by writing files to a designated directory
• The agents can immediately start using these self-created tools without any manual intervention

Dynamic System Prompt Updates

• Agents can update their own system prompts by reading from a persistent storage location (e.g. environment variable, S3 object)
• This allows the agents to continuously refine their knowledge and context

Learning from Interactions

• Strands agents can store conversation history and memories in a knowledge base (e.g. Bedrock, S3 vectors)
• They can then retrieve and leverage this past context to inform future responses

Meta-Agents and Orchestration

• Agents can dynamically create sub-agents with custom system prompts and tool sets
• These sub-agents can work in parallel or recursively, with the main agent orchestrating the workflow
• Concepts like "swarms" and "graphs" enable complex, self-organizing agent collaboration

Deployment to Agent Core

• Strands agents can be easily deployed to AWS Agent Core, a serverless runtime for autonomous agents
• This allows the agents to be scaled and run in a production environment with minimal overhead

Technical Details

• Strands SDK available in Python and TypeScript, with plans for more language support
• Agents leverage large language models (e.g. Claude, GPT-5) for their core capabilities
• Persistent storage options include Bedrock knowledge base, S3 vectors, and local file system (journal)
• Agent Core provides a serverless runtime with built-in memory and policy management

Business Impact and Use Cases

• Enables the creation of highly autonomous, self-improving AI assistants and research agents
• Reduces the engineering burden of defining complex workflows and error handling
• Allows for the emergence of novel, unpredictable behaviors that can adapt to changing needs
• Applicable in domains like personal productivity, scientific research, and open-ended problem-solving

Examples

• The presenters shared an example of a Strands agent that built its own weather calculator, text analyzer, and other tools during runtime
• Another agent was able to update its own system prompt and leverage past conversation history to provide personalized responses
• The meta-agent concept was demonstrated with agents dynamically creating sub-agents to perform parallel and recursive tasks, sharing a common context

Key Takeaways

• Strands empowers the development of truly autonomous, self-improving AI agents that can adapt and evolve beyond what their initial creators envisioned
• The framework addresses the challenge of engineering complex, error-prone workflows by offloading orchestration to the models themselves
• Deployment to Agent Core makes it easy to run these agents in a scalable, production-ready environment
• While offering significant benefits, the stochastic nature of these agents introduces new challenges around testing and validation that must be carefully considered