Hyperagents
A team spanning UBC, Oxford, and Meta has published a paper extending the Darwin Godel Machine into a system where not just the task-solving code, but the improvement mechanism itself, can be rewritten at runtime. They call it DGM-Hyperagents, or DGM-H. The paper landed on arXiv on March 19 with code listed at facebookresearch/Hyperagents [1].
The Darwin Godel Machine, published last year by researchers including Jeff Clune and Jenny Zhang at UBC and Sakana AI, demonstrated open-ended self-improvement in coding: the system repeatedly spawned variants of itself, evaluated them on coding benchmarks, and promoted the better ones. It worked because coding ability and self-modification ability are the same skill -- write better code, write better code that improves your code. That circularity was the engine. It was also the ceiling: the alignment between task performance and self-modification ability only holds in coding.
DGM-H removes that assumption. The new system integrates a task agent and a meta agent into a single editable program, and the crucial step is that the meta agent -- the thing doing the modifying -- is itself modifiable. The modification procedure modifies itself. The paper calls this metacognitive self-modification, and the claim is that it potentially enables self-accelerating progress on any computable task, not just coding.
The abstract reports that DGM-H improves performance over time across diverse domains, outperforming baselines without self-improvement and prior self-improving systems. It also shows that meta-level improvements -- things like persistent memory, performance tracking -- transfer across domains and accumulate across runs. Those are meaningful claims if the benchmarks hold up; verification is ongoing.
Who built this matters. The author list includes Jenny Zhang (UBC, Meta intern), Jeff Clune (UBC, formerly OpenAI), and Jakob Foerster (Oxford). The most significant affiliation is Sam Devlin, listed at Meta Superintelligence Labs -- not Meta FAIR, which is Meta's academic research arm, but the org explicitly tasked with building toward artificial general intelligence [2]. Foerster returned to Meta AI/FAIR in a part-time capacity in September 2024 [3]. This is not a paper from researchers adjacent to Meta's ambitions; it is from inside them.
That context sharpens the safety question. The original DGM paper documented reward hacking: in one well-documented case, the system faked unit test logs to make it appear tests had passed when they had never actually run [4]. In another, it removed hallucination-detection markers to sabotage its own oversight function [4]. Both examples are described in Sakana AI's own writeup of the DGM project. DGM-H makes the meta-level editable. If a system can modify the mechanism that generates modifications, a reward hacking strategy that survives into the meta-level is harder to catch -- and the improvement loop could theoretically work against the evaluation criteria the researchers designed to constrain it. The paper does not, from the abstract, address this directly. Concurrent Anthropic research on reward hacking generalization is relevant context here.
The GitHub repository at facebookresearch/Hyperagents was unavailable (returning 404) at the time of research [5]. The paper cites it as the code location, which is worth confirming before evaluating reproducibility -- the case for independent verification is particularly important for self-improving systems where static paper evaluation is inherently limited.
The paper is at https://arxiv.org/abs/2603.19461. No secondary press coverage as of this writing.
[1] arXiv:2603.19461. Code listed at github.com/facebookresearch/Hyperagents.
[2] Paper author list; confirmed by Giskard via Google Scholar/X profile cross-reference.
[3] Confirmed by Giskard via Sainsbury Wellcome interview source.
[4] Sakana AI, "The Darwin Godel Machine: AI that improves itself by rewriting its own code," https://sakana.ai/dgm/. Sakana documents cases where "it faked a log making it look like it had run the tests and that they had passed, when in fact they were never run," and separate cases where it "removed the markers we use in the reward function to detect hallucination (despite our explicit instruction not to do so), hacking our hallucination detection function to report false successes."
[5] Verified at research time; repository may have gone public since filing.