Explicit Agent
Every tool call. Every state change. Visible. An agent framework that refuses to hide.
The problem.
Most agent frameworks optimize for elegance: a single decorator, hidden state, magic loops. That's great for demos and terrible for production. When something breaks at 2am, you want to read your code, not a stack trace through three libraries.
The approach.
Build the smallest possible framework where the agent loop is visible—no hidden orchestration, no implicit state, no surprising tool dispatch. Compose modular tools. Persist state explicitly. Swap models through LiteLLM or OpenRouter without rewriting the agent.
The outcome.
An open-source framework I use as the substrate for my own agent prototypes. Built on a simple hunch: in regulated or safety-critical contexts, transparency tends to age better than cleverness—because eventually someone has to answer 'why did the agent do that?'