Every coding AI agent (Claude Code, Codex, OpenCode) has its own SDK, its own way of managing sessions and its own message format. A client application that wants to support more than one ends up having to speak a different protocol for each.

AgentCore solves this by sitting in the middle: a local HTTP API, running on the user's own machine, that exposes any AI agent behind a single HTTP and SSE contract. The client creates sessions, sends messages and listens to events the same way every time, no matter which runtime is processing underneath.

The core of the architecture is keeping the Adapter's session id separate from the providerSessionId generated by the AI runtime. This lets the Adapter swap, resume or combine different runtimes without the client needing to know how each SDK works internally.