Systems on Glass is an AI-automation side project of mine. Every episode is generated by Claude (Anthropic's AI, working through Claude Code in my own video studio): the script, the animations and 3D scenes, the edit, and the narration in an AI clone of my voice. I choose the topics, bring the stories from my own work, and review each episode before it goes up. Not affiliated with Anthropic.
I typed a few sentences into my terminal: add a drawing feature to my video studio, test it, and merge it. By evening there was new code, every test passing, and a merged pull request. A chatbot can't do that. So what is an agent, actually?
The short version
- Not a smarter chatbot. The model inside can be exactly the same. What changes is everything around it.
- Who runs the loop? With a chatbot, you do: copy the code, run it, paste back the error. With an agent, the model does.
- The loop: think about the goal, act by calling a tool, observe the real result, check. Not done? Go round again. Researchers named it ReAct in 2022: reason, then act.
- Tools are the hands. The model can't run anything itself. It asks for a tool, your code runs it and hands back the result. No tools, no agent.
- What stops it: a clear goal, a budget (steps or dollars), and a human for anything you can't undo.
until: done · or the budget runs out · or it needs a human
My favourite moment from building this series: the agent rendered a 3D scene, looked at the frame, spotted a ghost rectangle in one corner, and fixed the cause instead of hiding it. That's the point of the loop: judge the real result, not what you hoped for. And it asks me before every paid render.
Further reading
- Building effective AI agents — Anthropic (2024): workflows vs agents, and why agents are "typically just LLMs using tools based on environmental feedback in a loop".
- ReAct: Synergizing Reasoning and Acting in Language Models — Yao et al. (2022), the paper behind think-then-act.
- Writing effective tools for AI agents — Anthropic, on designing the "hands".
- LLM Powered Autonomous Agents — Lilian Weng's overview: planning, memory and tool use.
- Claude Code — the agent I used to build this series (docs).
- The Teammate in My Terminal — my post on the setup: rules, skills, allow lists, hooks, and keeping merge for myself.
Next episode: Caching, and how seventeen calls got cheap. All episodes →