Sample Agent Platform With Agentcore — Development skill for Claude Code

Agent Platform with Amazon Bedrock AgentCore

A reference implementation of an **internal agent platform** built on [Amazon Bedrock AgentCore](https://aws.amazon.com/bedrock/agentcore/). It shows how a platform team can offer two hosting models behind one portal:

• Interactive cloud workspaces — launch a full Claude Code CLI inside an AgentCore Runtime and use it from a browser web terminal. The process lives in a persistent tmux session, so disconnecting or switching sessions keeps the conversation and in-flight work running; files and conversation history persist to S3 and survive container restarts.
• Headless agent kernels — publish Claude Agent SDK based agents as AgentCore Runtime endpoints that any application can invoke through a single /invocations contract.

Both kernels route model traffic through a configurable **LLM gateway** (e.g. LiteLLM) with a **fixed egress IP** (VPC mode + NAT Gateway), so the platform works in enterprises that enforce model allow-lists, budgets and source-IP restrictions. The gateway key lives in a platform-side service and **never enters a session container** — a session's user is root in its own microVM, so a kernel gets a short-lived, per-session grant rather than a credential. Direct Bedrock access (cross-region inference) is supported as an alternative.



The hero diagram above is the map; two focused diagrams zoom into the parts that carry the security story: the [management / data-plane split](docs/images/data-plane-split.svg) (one backend image, two deployments — the console cannot front production traffic) and the [customer-owned MCP hub chains](docs/images/mcp-hub-chains.svg) (how production applications and the Dev Workbench reach a self