agentcore documentation

agentcore is an experimental Go-native runtime for coding agents that execute inside managed sandboxes. It provides the model/tool loop, conversation state, session lifecycle and typed event boundary without bundling a terminal UI or a JavaScript runtime.

The project is under active development. Its public API and wire behavior are not stable yet.

Start here

• Getting started covers building and running the CLI.
• Runtime lifecycle defines Agent, Session, Turn, Model call, Tool call and Subagent.
• Runtime architecture explains the current package and execution flow.
• Capability status records what is implemented and what is intentionally deferred.

Reference

• Configuration
• Events
• Tools
• Go API

Design principles

• Small runtime footprint. A native Go process rather than a desktop CLI stack inside every sandbox.
• Explicit semantics. Session, turn, model call, tool call and sub-agent have distinct lifecycle definitions.
• Embeddable core. The runtime communicates through Go types and channels; the outer process decides how events are transported.
• Conservative abstractions. Interfaces are introduced at proven I/O or ownership boundaries, not for hypothetical integrations.
• Correctness before breadth. Fewer capabilities with tests are preferred over broad but incomplete compatibility.

Scope

agentcore is not currently a full Claude Code replacement, a durable workflow engine or an agent framework with every provider and storage integration. The scope is deliberately narrow while runtime semantics and compatibility are being established.