Tools that make agentic systems easier to see, build and measure.
An agent is a model in a loop: step after step, the model proposes an action and the harness carries it out, until it stops. A chat box shows only the question and the answer. Our projects open up the loop in between, with models that run on your own machine, so you can watch each step instead of taking it on trust.
Based in Taiwan · Docs in English and Traditional Chinese
One turn of an agent: the model reads the context and proposes an action, the harness carries it out in the environment, and the result goes back into the context, until the model stops.
Projects
Open source, and built so you can run them on your own machine.
Watch an agent loop run step by step, down to every token the model writes: the token's probability, the candidates it beat, the special tokens, and the tool calls the harness reads out of them.
Gemma 4 E2B or Qwen3.5 0.8B runs on your machine with transformers, or on vLLM. You can also try MTP speculative decoding, attach images, audio or video, give the model a Python sandbox, or hand the loop to pi.
How it works: the 18 stages of one turn, each linked to the lines of code that do it
Experiments to try: sampling, stop conditions, human approval, MTP, images as tokens, code in a sandbox
More tools are on the way: one for prototyping a new agent system from an idea, and one for testing a change to an existing system against your own tasks and hardware. They'll be listed here once they're public.
Learning together
A study group that works through how agents are built, checked and shipped, notes on the AI engineering talks we learn from, and a playbook for learning as a team.
Agent Atlas strings seven courses from CMU, Stanford and Harvard into one learning path for AI agents, in 12 units. It starts with building an agent and checking what it can do, and goes on to shipping software and improving it over time.
The study group follows this path, meets online and keeps shared notes. Sessions are in Traditional Chinese. Agent Loop Visualizer is the hands-on companion to Unit 01.
Notes on conference and YouTube talks about AI engineering, sorted into 9 categories, from BI and the semantic layer to agent evaluation, context and memory, and model training. Each talk has a short summary of its key points, with its full notes a click away and a link to the video.
Nine insights draw on all the talks together. Select a sentence in a talk's notes to save it as your own note; it stays in your browser. The site is in English and Traditional Chinese.
A playbook for how a team of 3–4 engineers keeps turning what the team doesn't know yet into verified, shared knowledge. The questions come from real work, each lab answers one of them with a small experiment, and what the team learns becomes evals, tests and decisions.
Six mechanisms keep it going: a channel for new questions, a question backlog, a 75-minute lab every two weeks, failures turned into evals, an owner and a shadow who rotate each quarter, and a quarterly learning review. The page is in Traditional Chinese and English.
Agentic System Lab
Agent Loop Visualizer