In short
In today's roundup: tools for safer development, new approaches to evaluating AI agents, and hands-on experience from an autonomous product team.
In today's roundup: tools for safer development, new approaches to evaluating AI agents, and hands-on experience from an autonomous product team.
🔥 Hot:
🔹 [Canonical is developing a C-to-Rust code translator](https://www.opennet.ru/opennews/art.shtml?num=66099) — The tool is aimed at large repositories and should translate them into a Rust representation with no unsafe operations. 🔹 [A personalised mRNA cancer vaccine has passed phase three trials](https://habr.com/ru/articles/1072308/?utm_campaign=1072308&utm_source=habrahabr&utm_medium=rss) — Merck and Moderna tested the vaccine in melanoma patients; the piece examines AI's role in developing the approach.
➡️ Useful reading:
🔹 [FinSkillBench has been released for evaluating AI agents in investment management](https://arxiv.org/abs/2608.18099) — The benchmark tests not only the plausibility of text but work with data, calculations, specialised methods and auditable results. 🔹 [Researchers propose treating multi-agent system problems as concurrent access control](https://arxiv.org/abs/2608.18092) — Parallel reads and writes of shared state lead to stale data, lost updates and contradictions. 🔹 [The authors propose testing AI agents by behaviour, not only by result](https://arxiv.org/abs/2608.18081) — For systems that act in a changing environment and adapt over time, metrics of final effectiveness alone are not enough. 🔹 [A study warns that today's model cards are not enough to govern open-weight models](https://arxiv.org/abs/2608.18086) — The authors point to gaps in informing developers and users about model risks after publication.
➡️ Discussions and cases:
🔹 [An author has assembled an autonomous system with an AI product manager and a team of agents](https://habr.com/ru/articles/1072310/?utm_campaign=1072310&utm_source=habrahabr&utm_medium=rss) — The system closes product backlog tasks around the clock and in parallel; the author shares the results and how the approach is built.
📝 If you would like to add other news and materials to the list, write in the comments.