AgentCore: A Minimal Go Library for Building AI Agent Applications

AgentCore is a Go library for developing AI agent applications. It is designed with a minimal, composable core to allow for open extensibility, aiming for reliability over a complex, all-in-one approach. The library's architecture features a standalone loop and a stateful Agent, with an event stream driving UI updates and a context layer managing prompt projection and overflow recovery.

Key features include support for both single and multi-agent setups. For single agents, it offers tool call gating, allowing users to define permission policies. Multi-agent capabilities are facilitated through a SubAgent tool, which supports four execution modes: single, parallel, chain, and background. The library also provides mechanisms for steering and injecting messages into agent runs, as well as a unified event stream for all lifecycle events.

AgentCore includes built-in tools for file operations (read, write, edit) and supports swappable models and custom LLM adapters. Its context management system automatically summarizes conversation history to manage context window limits, keeping recent messages and summarizing older ones into structured checkpoints.