Stanford's Self-Improving Agent Course: What's Actually Being Improved Is Often Not the Model

Omar Sar posted Stanford CS329A 'Self-Improving AI Agents' course playlist on X, and the notes are still being organized. The two most common questions in the comments were: how does an agent know it has really gotten stronger, and how does it avoid learning its mistakes?

These two questions are exactly where the course is most worth watching, and exactly where this term is easiest to hand-wave.

But among what's been released this time, what's more worth opening first is not the course, but the two paper collections Omar Sar posted. Harness Engineering has collected 829 papers, and Self-Improving Agents has collected 545, updated daily. You can connect them to your own agent through MCP and talk directly to the papers. The account is free.

Omar Sar himself put it bluntly in a reply: good references can save everyone a pile of conversations in which people talk past each other. That's not polite, but it's accurate. Over the past two years, discussion of 'self-improving agents' has gotten increasingly messy. The same term can mean the model changing its own weights, the agent changing prompts, the agent writing experience into memory, or the agent rewriting its own scaffolding. Everyone is talking about something different.

Put the 829 Harness Engineering papers together and a judgment emerges: progress mainly comes from the shell outside the model, not from the weights themselves.

In AI circles, harness refers to the whole system around a model: the agent loop, context management, tools, memory, orchestration, verification, and runtime environment all count. With the same model weights, putting on a different harness can make results differ by a wide margin. The Prime Agent numbers in the list are repeatedly cited: under exactly the same weights, ARC-AGI-3 went from 30% to 95.5%, without a single word of the weights changing.

Similar research keeps piling up. One paper is literally called 'The Harness Is the Model,' making the same point. Another, 'Harness or Model?,' uses a contamination-controlled private test set to separate 'the gain from scaffolding' from 'the gain from the model itself.' The conclusions broadly agree: many improvements attributed to 'a stronger model' are actually the result of fixing the scaffolding.

This is also why the list starts with GPT-2. The GPT-2 generation had no tool calls, no skills, no memory; just a while-not-EOS loop plus top-p sampling. Over six years, the repeated task has been to stuff new actions it is allowed to take into that loop. Tool calls, memory read/write, skill accumulation, sub-agent addressing, recursive REPL—each step adds a new class of permissions to the loop. From this perspective, so-called self-improvement is mostly giving itself more permissions, not giving itself a new brain.

Follow this line of thinking and you run into a problem no one can avoid: verification.

A reply on X put this most clearly: the life or death of the self-improvement loop depends entirely on that scorer. If the agent can see the judge that evaluates it, it will learn that judge, and learning how to do the task well becomes secondary. The only thing that can prove it has really gotten stronger is a frozen held-out task set that it cannot see.

Another reply supplied the other half of the same problem: self-improvement only holds when someone is responsible for 'what counts as better'; otherwise the agent will optimize the metric that is easiest to move.

These two points are really two sides of the same thing. Once the scoring mechanism is visible to the agent, the agent will optimize it in turn. That's why so many papers in the list talk about terms like acceptance gate, statistical guarantee, and masked erosion. They sound ugly, but they are all engineering answers to 'how do you know it has really gotten better?'

0xSoldier's question is quite representative: I really want to know about eval—how does an agent know it has really gotten stronger, rather than grading its own homework? Varun Chopra asked the other half: how does it avoid locking in its own mistakes? Under both questions, the paper titles posted get longer and longer, and no answer can be dismissed in one sentence.

Back to the resources themselves. Both paper collections have reading paths laid out: 21 papers on the Harness side and 31 on the Self-Improving side, so you don't have to dig through 829 yourself. The MCP tool can be connected to an agent so that, during research, the agent pulls these papers itself rather than relying on a fixed reading list. Omar Sar emphasized this twice: it's all free, and you can use it just by registering an account.

From a usage perspective, the more reliable path is to first look at those few dozen curated lists, see how others divide generations and failure modes, and then decide whether to dive into the full 545. The 829 collection leans engineering, suitable for people already building agents; the 545 collection leans toward the methodology genealogy, suitable for people who want to understand what this term actually means.

As for the course itself. The notes aren't out yet, but the playlist is out first. Someone in the comments joked that the playlist arrived before the notes, so the notes had better live up to it. Stanford's course belongs to the category you can read as a textbook, not the slide-deck version of a launch event.

But the real question is not in the course, nor in how many papers are on the list. The real question is: when an agent says it has gotten better, do you believe it? That answer, for now, is still not in the model's weights.