Google Research published a tutorial implementing RRSI (Regularized Recursive Self-Improvement) for LLM agents to rewrite their own harness while keeping the underlying model frozen
Research publication Provisional 72% confidence first seen
Two articles describe Google Research releasing a practical guide/tutorial for RRSI, an approach where an LLM agent can revise its own prompting, tools, memory, control flow, and sub-agents while the base model remains fixed. The tutorial turns the paper’s edit-selection logic into runnable code, and reports evaluation details alongside results showing improved performance on SWE-bench and Polyglot in self-rewriting coding-agent experiments.
Google Research
Google Research RRSI Guide: Mastering Self-Improving AI Agents
RRSI (Regularized Recursive Self-Improvement)
Sakana
Jeff Clune
LLM agent
MarkTechPost
Polyglot
Python
SWE-bench