Abstract syntax tree

So the harness is treated as a stateful program that dictates what the model sees and how it acts.

By letting an agent optimize the AST, the abstract syntax tree of the harness itself, the researchers saw a 7.7-point improvement over state-of-the-art baselines on complex text classification.

All while using four times fewer context tokens.

So if you're listening, stop worrying about migrating your entire stack to the newest model API every three weeks.

Exactly.

When an agent writes code, you don't ask it if the code is correct.

You run it through an AST parser to programmatically guarantee syntactical correctness before it hits the orchestrator.

You check the actual database rows.

The red squiggles.

It parses your code into an abstract syntax tree, an AST.

So it actually understands the structure and the logic of the code, not just the raw text strings?

You nailed it.

Okay, so how does plan mode alter that transaction boundary?

Plan mode enforces a read-only contextual discovery phase.

The agent still retains full access to read the file system, analyze the AST via the language server, and query the MCP tools.

So it can still see everything.

And you can perform at a higher level than you could just on your own.

There are areas of programming I know nothing about, like using ASTs, abstract syntax trees across multiple different languages.

These are APIs with hundreds of possible operations.

No human being is going to wrap their brain around the full API in one of these AST analyzers, much less five different ones.

And you can perform at a higher level than you could just on your own.

There are areas of programming I know nothing about, like using ASTs, abstract syntax trees across multiple different languages.

These are APIs with hundreds of possible operations.

No human being is going to wrap their brain around the full API in one of these AST analyzers, much less five different ones.

Why? I could explain.

Because in source code, we have this concept of AST3.

So whenever you compile the language, this compiler, like any one, all the languages use the same approach.

It should understand what is the function, split them across, which function called another function, which arguments to properly compile that.

This function doing that because it's called by this immediately.

That's useless.

Like implementing GraphQL instead of AST3 is harmful.

It's making it worse.

Hold on.

So it literally rewrites the AST.

If I chain those three abstract functions, the compiler doesn't execute them sequentially.

Nope.

Tell me about it.

Yeah, so I've been working on a code mod, which is a script that analyzes the AST of your project.

And throughout AST analysis, it can look for patterns.

It can detect certain things, certain API usages.

And it can either just report on them, or it can make deterministic changes to your AST.

this is pretty valuable because uh like in in today's age you can you probably think that you can hand your project and a migration guide to an ai and let it do all the work which Yeah, and most of the time, that's going to work.

That's probably going to work, especially if you have a small to medium project.

Some of them you actually have to think about them in terms of like how to restructure your code, how to move some some logic around so that it actually follows the pneumatic patterns.