Amazon has introduced Bedrock AgentCore, a new runtime environment designed to host coding agents securely without the need for a laptop. Developers often keep their laptops open to maintain coding agent sessions, but this practice poses security risks. AgentCore provides a dedicated environment for each session, ensuring isolation and persistent workspaces. The solution addresses the limitations of running agents on laptops, where shared resources and open sessions can lead to security vulnerabilities and data exposure. The new runtime is part of Amazon Bedrock, which already supports models like Claude and OpenAI's offerings.
AgentCore creates an isolated Linux microVM for each session, complete with a persistent workspace, a real shell, and deterministic command execution. This setup ensures that each agent operates in its own environment, avoiding conflicts with other processes. The system also includes an identity layer that allows agents to act as the user who triggered them, along with a gateway that provides access to external tools through a unified Model Context Protocol (MCP) endpoint. Observability features like Amazon CloudWatch are integrated to track every action the agent takes.
The article highlights the security risks of running coding agents on laptops, where shared resources like the shell, filesystem, and credentials can expose sensitive data. It notes that agents often share the same machine, leading to conflicts over resources such as Postgres, SSH keys, and network access. These issues make laptops unsuitable for running multiple agents in parallel. The solution proposed by AgentCore is to use a dedicated machine per agent, allowing for secure, isolated environments that can survive reboots or extended periods of inactivity. Source: awsml
