A new arXiv paper presents Terminal-RSI, a method for building recursive self-improving terminal agents. The authors frame the core problem as two linked difficulties: turning verifier outcomes into usable policy updates, and converting partial progress into full task completion.
Terminal-RSI alternates between Recursive Self-Rewrite (RSR) and policy refinement. The abstract suggests this alternating loop is what lets the agent improve from its own verified behavior, though the full mechanism and evaluation details are not included in the abstract.
Since this is a single source, there are no independent results to compare against. The paper is announced as new on arXiv, and the claims here reflect only what the abstract states.