Show HN: AgentSilex – Transparent agent framework you can read in 30 minutes (github.com)

AgentSilex is a new minimalist, transparency-first agent framework aimed at developers, educators, researchers, and companies who want a readable, hackable foundation for building AI agents. The core implementation is intentionally small — roughly 300 lines — so you can audit, modify, and extend every part rather than wrestling with a black‑box system. It’s built on LiteLLM and advertises seamless model switching across 100+ providers (OpenAI, Anthropic, Google Gemini, Mistral, Azure, local LLMs, etc.) with a single configuration change, and installs with a simple pip command. Technically, AgentSilex provides an intuitive API (Agent, Runner, Session, @tool) for tool calling, session-based conversation history, streaming responses, async and parallel tool execution, and agent-to-agent handoffs (specialist sub-agents and agents-as-tools). It auto-extracts parameter schemas from Python type hints, supports Model Context Protocol (MCP) clients, and includes built‑in error handling, state persistence, and human‑in‑the‑loop approval flows. Observability is first‑class via OpenTelemetry tracing compatible with Phoenix (OTLP endpoint) so you can visualize tool calls, handoffs, and traces. The result is a practical, educational platform for rapid prototyping and experimentation where you can both learn agent internals and deploy customizable production workflows.