Z.ai appears to be pushing harder on recursive self-improvement. An intern’s public profile we spotted describes a goal of “100% RSI”, with the loop running on coding agents first and now extending to medical agents. That percentage is an ambition in the profile, with no published result attached. 1 2
The sequence is the interesting part. Coding is where the profile says the loop started; medical agents are the next area it names. This gives us a glimpse of a broader effort to improve agents through the work they do.
In our post, we described promising but unfinished work on coding agents that produce their own training data. That is the part we are watching: whether useful coding experience can feed back into training and produce a better agent. 1
The same profile names GLM-5.5 in work on the next-generation base model and describes a separate research topic on self-evolution in financial coding. Our GLM-5.5 report covers that model reference and the internal post-training details.
It is one reason we are expecting GLM-5.5 to have a strong coding focus. We still need a released model and results to see how much of this work shows up in practice.
The linked GitHub commit is no longer publicly retrievable, so we have kept the original capture above.
Sources & references
- Lumina: Z.ai’s work toward recursive self-improvement ↗
31 October 2026. Our original report, verified through X’s public oEmbed response. It describes promising but unfinished work on coding agents producing their own training data; expectations for GLM-5.5 coding performance are our interpretation.
- Captured researcher-profile commit ↗
Personal-profile commit supplied with our original captures. GitHub’s commit page, patch endpoint and API returned 404 when checked on 31 October 2026. The excerpt is retained in the unchanged supplied screenshot; it is a researcher’s account of their work, not a Z.ai release announcement.
