Tiny Software Factories: Agentic Engineering DONE WRONG

In short

The video introduces Tiny Software Factories: durable, proactive agentic systems built for one specific problem, with humans near the loop rather than repeatedly prompting in it. It demonstrates an intentionally imperfect architecture that ships quickly through features like agent access parity, customizable controls, and scheduled cron prompts, while contrasting it with more autonomous out-of-loop systems.

Key points

• A Tiny Software Factory uses one proactive agent in a durable environment to work around the clock on one problem.
• Near-loop operation shifts the human role from prompting to giving feedback, making decisions, and approving completed work.
• The factory starts as an empty template and adds an always-on agent, product lists, generative pages, and a central focus view.
• Agent access parity gives the agent a matching tool for every UI control, including pages, settings, and coding actions.
• The Core Four are context, model, prompt, and tools; the example uses 200 KB auto-compaction and 500 KB forced compaction.
• An extensible harness combines factory, coding, pages, image-generation, and factory-control capability slices.
• Cron prompts contain a name, schedule, and prompt, allowing repeatable work to continue even when the browser is closed.

Written by an AI model from the video’s captions and description.