Labs·English · Vietnamese

How I built Soot, explained properly.

Labs is where I write down what I actually built, for someone who has never built an agent — no jargon left standing in for an idea, and every mechanism traced to the code. Agent 101 is the course. The Labs below it are the ones that stand alone.

The Agent 101 series

Read in order; each lesson can stand alone, but the second one assumes the first.

  1. Lesson25 min readHow an AI agent worksWhat an agent is made of, built one piece at a time: a model that guesses words, a loop that keeps going, tools, a desk called context, and the rules that stop it.The LLMThe loopTools and skills+2
  2. Lesson22 min readContext, and what I do when it fillsThe second lesson: a Soot works for hours in a window that never grows. Here is every technique I built to keep it going — an unbroken prefix, lazy tools, clearing, spilling, compaction, checkpoints, subagents and memory — with the blocks that show each one.The context windowPrompt cacheLazy tools and skills+6

Other Labs

Standalone pieces, not part of the course.

  1. Lab18 min readAutonomous self-evolutionHow a Soot gets better from its own traces, why every change must beat cases it never saw, and how you undo any of it.Traces as evidenceTests from real runsHeld-out cases+2

How these pages are made

  1. 01First person. I built this, so I describe it as I built it — including the parts that cost more than I would like.
  2. 02Every mechanism has a file behind it. The list is at the bottom of each lesson, and anything I could not trace to the code or the design notes is not written there.
  3. 03Examples are simplified illustrations. Where numbers are real measurements, they are labelled as measurements of one run, never as benchmarks.
  4. 04Interactive blocks are illustrations, not a live agent. They are built to show one mechanism at a time.

That is the whole idea

An agent is a loop around a model that has no memory. Everything else — tools, context, memory, checkpoints — is what makes that loop useful for a whole afternoon instead of one request.