Skip to content
Back to projects

May 2026

Agent SDK

Async-first SDK for building LLM agents: swappable providers, SQLite/ChromaDB memory, conversational/ReAct/autonomous modes, events and multi-agent.

  • Python
  • asyncio
  • Pydantic
  • Anthropic
  • OpenAI
  • ChromaDB
AI2026

Agent SDK is a Python library for building LLM agents without locking yourself into a provider, a memory backend or a runtime. Every piece is a swappable interface.

What it includes

  • Anthropic, OpenAI, Ollama and any OpenAI-compatible API (LM Studio, vLLM…).
  • SQLite (FTS5) memory, file-based memory or ChromaDB vector memory.
  • Three runtime modes: conversational, ReAct (reason → act → observe) and autonomous.
  • An observable event system and plugins that react to it.
  • Multi-agent: any agent can be wrapped as a tool with SubAgentTool.
  • Async-first, with streaming and tests.

Why

Agent frameworks often hide the loop, context and tools behind a rigid API. Here the configuration is explicit: you choose the provider, context strategy, memory and tools; the SDK handles the rest. It is the base I use to experiment with agents and the MCP protocol.

Back to projects