yaxin-luo/blog
EN/中
◇Page·Blog3postsHistory

Yaxin’s Blog

Research notes on post-training agents for long-horizon work: the model, and the harness around it.

  1. 2026-10-02Essayplaceholder

    What the harness knows that the model doesn’t

    Off the map2 min
  2. 2026-09-18Noteplaceholder

    How I read a long agent trace

    Off the map1 min
  3. 2026-08-30Noteplaceholder

    Notes on native multimodal models

    Off the map1 min
preview / what-the-harness-knows.mdhover to switch · click to read ↵
01 · what-the-harness-knows541 words · 2 min
Off the mapEssay2 Oct 2026placeholder

What the harness knows that the model doesn’t

A long-running agent is mostly scaffolding. Which parts of that scaffolding could a model learn to carry itself?

Inside

Model memory discipline judgment MemoryDisciplineJudgmentModel task lists, scratch files, summarieswhen to stop, verify, askrouting, “good enough” callswhat the weights already know
Figure 1
Figure 2chart
Figure 3chart

Outline

  1. 01Three kinds of scaffolding
  2. 02Pricing a rule
  3. 03What removal might look like
  4. 04A tiny example
  5. 05Which rules move first
EN · 中Read · 2 min