AWS Releases Tools for Safer AI Agent Development

AWS has released two open-source tools: a small model for agent decision-making and a secure sandbox for running agent-generated code.

Amazon Web Services has released two new open-source projects aimed at making AI agents more practical and secure, according to a report from InfoQ. The first, Strands Decider 2B, is a small language model designed to make quick, efficient decisions within an agent's workflow. The second, Strands Box, is a secure sandbox environment built to safely execute code generated by an AI agent. Together, these tools address two of the biggest hurdles in building autonomous agents: making them both cost-effective and safe when they need to interact with external tools, execute code, or manipulate data.

Strands Decider 2B tackles the problem of efficiency. Many AI agents rely on large, powerful models for every single step of their reasoning process, which is often overkill for simple choices like deciding whether to use a calculator or a web search tool. Using a massive model for such tasks is slow and expensive. Strands Decider is a specialized, 2-billion-parameter model trained specifically for this decision-making role, allowing the agent to operate much more quickly and cheaply. On the security front, Strands Box confronts the immense risk of running AI-generated code. Agents often need to write and run scripts to complete tasks, but this code is inherently untrusted. Strands Box provides an isolated environment that walls off the execution, preventing the code from accessing the host machine's file system, network, or sensitive processes, creating a safe playground for the agent to work in.

The release of these tools fits into a broader industry trend focused on moving AI from simple chatbots to autonomous agents that can take meaningful action. Companies from OpenAI to Microsoft are heavily invested in frameworks that allow AI to perform complex, multi-step tasks. However, the path to reliable and safe agents has been challenging, with early efforts often struggling with flawed reasoning and major security vulnerabilities. AWS's decision to release Strands as modular, open-source components reflects a maturation of the field. Instead of offering a single, closed agent product, AWS is providing the essential “picks and shovels,” encouraging developers to build on its ecosystem and driving adoption of its other cloud services in the process. This is a strategic move to become a foundational layer for the emerging agent economy.

For developers and technology leaders, these tools offer a tangible path to improving AI agent projects. Teams can now integrate a purpose-built decision model to cut down on latency and operational costs, while the security sandbox provides a concrete answer to difficult questions from security teams about running untrusted code. This modular approach allows for more control and transparency than relying on a single black-box API. Looking ahead, we can expect AWS to integrate these Strands components more tightly with services like Amazon Bedrock and Lambda. The success of these projects will ultimately depend on community adoption and how well they are integrated with other popular open-source AI frameworks. The key trend to watch is the industry-wide push to build agents that are not just capable, but also efficient, secure, and governable enough for real-world business applications.

For developers building AI agents, these tools address two critical bottlenecks: the high cost of using large models for simple decisions, and the major security risk of executing agent-generated code. AWS is providing open-source, modular solutions to make agents more efficient and safer to deploy.

By open-sourcing these tools, AWS lowers the barrier for companies to develop sophisticated AI agents, potentially reducing reliance on expensive models. Providing a secure sandbox also mitigates significant security risks, making it more feasible for businesses to deploy agents that interact with internal systems.