AWS Bedrock AgentCore Secures Coding Agents with Production Hosting

Amazon Web ServicesAmazon Web Services

AWS has updated Amazon Bedrock AgentCore to provide production-grade infrastructure for hosting coding agents. This offers isolated environments, centralized credentials, and audit logs, moving agent execution from insecure local setups to a managed cloud environment. The platform addresses common developer challenges with untracked usage and inconsistent access controls for tools like Claude Code and Codex.

AWS has updated Amazon Bedrock AgentCore to offer production-grade infrastructure for hosting coding agents. This new capability provides isolated environments, centralized credential management, audit logs, and platform-enforced security for tools such as Claude Code, Codex, and Kiro. It moves agent execution from local machines to a managed cloud environment.
Dedicated environment
Isolated Linux microVM
Session storage
Persistent /mnt/workspace (14 days inactivity)
Interactive shells
PTY-backed terminal access
Max runtime
Up to 8 hours per session
Supported agents
Claude Code, Codex, Kiro, OpenCode, Cursor CLI, Gemini CLI
Max mounts
5 filesystems (S3 Files, EFS) per runtime

Developers often run coding agents locally, storing sensitive tokens in .env files, which leads to untracked usage and inconsistent access controls. This update addresses those security and operational challenges by providing dedicated microVMs for each agent session. It enables parallel execution of multiple agents without resource collisions and ensures work persists across reboots or network disconnections.

Amazon Bedrock AgentCore allows you to host any coding agent, packaging it as a container or zip file, and selecting models from Amazon Bedrock or external APIs. The platform offers persistent storage for agent workspaces, interactive shell access into running microVMs, and secure tool access via AgentCore Gateway and Identity. This builds on previous managed agent harness capabilities for AgentCore, enabling parallel execution for tasks like fixing GitHub issues or benchmarking model performance.

View the full update on aws.amazon.com
AWS AI
AWS AI
@AWSAI
X

Developers are already running coding agents in Kiro, Claude Code, and Codex. Many are running setups with tokens in .env files, leading to untracked usage and inconsistent access controls. Amazon Bedrock AgentCore gives you one place to host any coding agent with production-grade infrastructure: isolated environments, centralized credentials, audit logs, and security enforced at the platform layer. https://t.co/Nn4WWKUSr8

View on X

Still wondering? A few quick answers below.

Amazon Bedrock AgentCore addresses the challenges of running coding agents locally, where developers often use insecure setups with tokens in .env files. This can lead to untracked usage, inconsistent access controls, and security risks like prompt injection. The platform provides a secure, managed environment for these agents.

AgentCore provides isolated Linux microVMs for each agent session, ensuring that agents do not collide on shared resources like local databases or SSH keys. This physical isolation enhances security and allows multiple agents to run in parallel without interference.

Yes, AgentCore Runtime is model agnostic. You can route your coding agent's model calls through Amazon Bedrock, which hosts models from Anthropic, OpenAI, and others, or directly via providers' APIs, or through your own LLM gateway.

AgentCore Gateway manages tool access, exposing a single Model Context Protocol (MCP) endpoint for tools like GitHub and Jira. AgentCore Identity securely holds credentials in AWS Secrets Manager and a Token Vault, attaching the right downstream credential for each tool call without exposing secrets to the agent's environment.

If an agent session remains idle past its configurable timeout (15 minutes by default), the compute resources for the microVM shut down. However, the persistent /mnt/workspace filesystem, along with any S3 Files or EFS mounts, remains intact. You can resume the same session ID later, and a fresh microVM will mount the same files.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update