Gödel Agent

Gödel Agent is a self-referential agent framework introduced by Yin, Wang, Yang, Mao, and Chen (2024). It uses an LLM to modify its own agent code, prompts, and tool set, then empirically validates the new version against a benchmark. The framework is named for the theoretical inspiration from Schmidhuber's Gödel machine (2006), which requires a formal proof-of-improvement step before accepting any self-modification. Gödel Agent uses empirical benchmark improvement in place of a formal proof, which makes it more practical but less guaranteed.

The framework operates in a loop. The LLM proposes a self-modification (rewriting a function, adjusting a prompt, adding a tool), the modified agent runs against a benchmark, the result is fed back, and the modification is accepted if performance improves. The system maintains a history of modifications and can explore alternative agent designs through parallel proposal streams.

Gödel Agent scored zero on ARC-AGI-2 in the Meta^n comparison. The authors of Meta^n interpret this as evidence that single-layer self-modification caps at roughly meta-depth 2, because beyond that point, the system destabilizes: each new layer must leave the previous editing machinery untouched. The framework has been extended in Polaris (arXiv 2603.23129, March 2026), which adds experience-abstracted prompts to make the framework practical for small language models.

Gödel Agent is adjacent to dgm-darwin-godel-machine (a 2025 follow-up that adds open-ended evolutionary search), meta-n-recursive-self-improvement (a 2026 follow-up that achieves meta-depth beyond 2 by fixing the meta-operation), and recursive-self-improvement (the theoretical category). It differs from these in being a single-layer system with exactly one self-modifying step per cycle, while the more recent meta-n is multi-layer by construction.