運作說明 · How it works · Agent Loop 可視化

Agent Loop Visualizer displays how language models execute agentic workflows step-by-step. A chat interface on the left mirrors the agent's reasoning loop shown on the right: the model reads context, proposes actions, a harness executes them, and results feed back until completion. Models (Gemma 4 E2B or Qwen 3.5) run locally in your browser via Hugging Face transformers, exposing token probabilities, candidate tokens, and generation speed. MTP speculative decoding accelerates Gemma 4. The companion "How it works" guide maps each step to source code with file names and line numbers. Bilingual interface in English and Traditional Chinese.