News · 2026-08-12

A self-improving coding agent that compares notes with a rival lineage

Self-improving coding agents rewrite their own source code based on a single failed run, which means they are guessing at which part of themselves caused the failure. Changzhi Liu, Yilun Liu, Sikuan Yan, Volker Tresp, and Yunpu Ma propose using the archive instead, adding two edit operations that draw on multiple runs and on a competing agent's evidence. They call the result the Mendel Gödel Machine, and it improves on the single-trajectory baseline across both a short coding suite and a repository-level one.

Key facts

• Published as the Mendel Gödel Machine, with a project page and public code.
• Adds two self-modification operators to the standard clonal mutation: reaction-norm mutation and cross-lineage hybridization.
• Neither new operator requires additional task evaluations, since both reuse trajectories already in the archive.
• Evaluated on SWE-bench and Polyglot, with gains that transfer to different backbone models.

The diagnosis problem

An agent that edits its own scaffold has to answer a question every debugger knows: given that something went wrong, what exactly caused it. One failed run is thin evidence. Many possible edits are consistent with it, and most of them are wrong. The agent picks one, applies it, and either gets lucky or introduces a new problem while leaving the original.

Meanwhile the agent is sitting on an archive of every attempt it has ever made, which contains precisely the comparative information that would narrow the search. The paper's framing is that existing methods "derive self-modification from a single failure trajectory at a time, overlooking rich comparative signals available in the agent's expanding archive of past attempts."

The Mendelian framing in the name is about controlled inheritance, and it is more than a metaphor. Mendel's insight was that you learn about a trait by comparing across controlled crosses, not by staring at one organism.

The two new operators

Reaction-norm mutation edits an agent based on its trajectories across multiple tasks at once. The term is borrowed from biology, where a reaction norm describes how one genotype expresses differently across environments. Applied here: run the same agent on several tasks and look at the pattern. A behavior that fails everywhere is a different kind of defect from one that fails only under specific conditions, and a single trajectory cannot distinguish them. Seeing the spread tells you whether you are looking at a broken component or a bad fit to one situation.

Cross-lineage hybridization edits an agent using the trajectory of a reference agent from a different lineage on the same task. This is the controlled experiment: same problem, two different approaches, one succeeded and one did not. Everything shared between them is ruled out as the cause, and what remains is a much smaller candidate set. It is the same logic as a diff, applied to behavior rather than source.

Crucially, neither operator requires new task evaluations. They reuse trajectories the system already has. Evaluation is the expensive part of this loop, so making the edits smarter without making them costlier is the substantive contribution.

What they show

The paper proves, under an additive fitness landscape model, that comparative evidence should improve the probability of editing the right thing, then demonstrates the same effect in controlled surrogate simulation before running real experiments. That order matters. It is a claim about why the method works, tested in a setting where the ground truth is known, before being tested where it is not.

On SWE-bench and Polyglot the improvements over the single-trajectory baseline are consistent, and they transfer: the evolved scaffold keeps its advantage on harder held-out benchmarks and when moved to different backbone models. The authors frame that transfer as evidence the system discovered reusable workflow-level improvements rather than benchmark-specific hacks, which is the distinction that separates genuine self-improvement from reward hacking.

That framing is not automatically true and it is the right thing to be suspicious of, but transfer to a different backbone is a real test. A benchmark-specific hack does not survive a model swap.

Where it sits

This lands in a busy lane. We have covered models that rewrite their own harness and gain 16 points while flunking office work, a model that improved itself by training only where it disagreed with itself, and an agent that edited its own runtime for 161 days. The through-line across all of them is that the bottleneck is not the ability to make an edit. It is knowing which edit to make, and knowing whether the edit helped.

The Mendel Gödel Machine attacks the first half. The second half, who checks the self-editing agent, remains open, and it is the harder one. Our notes on recursive self-improvement and self-play cover why comparative signal is such a recurring theme in systems that improve without a human in the loop.

The caveat

The theoretical guarantee holds under an additive fitness landscape, meaning defects contribute independently to performance. Real agent scaffolds are not additive; components interact, and a fix in one place changes what a fix elsewhere does. The proof is a reason to expect the method to help, not a guarantee it will. The gains are also reported by the authors on benchmarks they selected, with the usual caveat that self-improvement results are unusually sensitive to how much headroom the starting scaffold had. The code being public is what makes this checkable, and it is.

Primary source, verified: read the paper → (arXiv 2608.07645)

Ground Truth. (2026, August 12). A self-improving coding agent that compares notes with a rival lineage. Ground Truth. https://groundtruth.day/news/a-self-improving-coding-agent-that-compares-notes-with-a-rival-lineage.html

@misc{groundtruth:a-self-improving-coding-agent-that-compares-notes-with-a-rival-lineage,
  title  = {A self-improving coding agent that compares notes with a rival lineage},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/a-self-improving-coding-agent-that-compares-notes-with-a-rival-lineage.html}
}

Topics: agents · recursive-self-improvement · coding-agents · evolution · research