TECHNICAL REPORT / 2026

Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution

The report describes how Ouroboros changes its own runtime through reviewed Git history, preserves identity and memory across releases, and turns experience from ordinary work into structural improvements.

Anton Razzhigaev, Andrei Gritsaev, Andrei Kaznacheev, Nikita Dragunov, Roman Yampolskiy, and Andrei Kuznetsov.

What the report covers

• Reviewed core evolution across code, prompts, tools, context assembly, architecture, and dependencies.
• Hope, a 161-day deployment in which ordinary work and human interaction continually expose new problems and improvement opportunities.
• Benchmark campaigns on Terminal-Bench 2.1, OSWorld-Verified, CL-Bench, SWE-bench Pro, and GAIA, including limitations and audit corrections.
• The operational controls used to keep self-modification attributable, reviewable, and recoverable.

Paper and evidence

• arXiv 2608.08311 is the canonical preprint record.
• Hugging Face Papers carries the community discussion and linked artifacts.
• Benchmark evidence collection groups the public datasets and explorer.
• Benchmark Explorer provides a no-key view of the released evidence.
• Source repository contains the implementation, history, benchmark adapters, and methodology.

Citation

@techreport{razzhigaev2026ouroboros,
  title        = {Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution},
  author       = {Anton Razzhigaev and Andrei Gritsaev and Andrei Kaznacheev and Nikita Dragunov and Roman Yampolskiy and Andrei Kuznetsov},
  year         = {2026},
  eprint       = {2608.08311},
  archivePrefix = {arXiv},
  primaryClass = {cs.SE},
  doi          = {10.48550/arXiv.2608.08311},
  url          = {https://arxiv.org/abs/2608.08311}
}