Can a single agent improve itself by editing its own code?

The paper introduces a "self-improving coding agent" (SICA), "equipped with basic coding tools," that "can autonomously edit itself, and thereby improve its performance on benchmark tasks." The measured result: "performance gains from 17% to 53% on a random subset of SWE Bench Verified, with additional performance gains on LiveCodeBench, as well as synthetically generated agent benchmarks." The architectural move the authors stress is eliminating a split kept by the prior ADAS system, where "the target-agent... performs the task, and the meta-agent... improves the target agent." SICA instead "eliminates the distinction between meta-agent and target agent," operating "over the agent's full Python codebase" rather than ADAS's single forward function in a domain-specific language.

The mechanism is an archive-and-select loop: SICA keeps "an archive of previous agents and their benchmark results," and at each step takes the best-performing agent from that archive as the meta-agent, which "look[s] through the archive, identif[ies] an improvement, and implement[s] it"; the new agent is then evaluated on the benchmark set and stored. "Best" is defined by a fixed utility function combining benchmark score, wall-clock time, and dollar cost. Crucially, the editing is confined to the scaffold — tools, sub-agents, prompts, oversight code — not the underlying model: "our self-improving coding agent shows improvement in certain agentic tasks... which do not modify the model weights." The paper argues this absence of weight updates is itself a safety mitigation: without weight updates it is harder for the model to "encode e.g. stenographic messages in the chain-of-thought," so observability (an interactive web interface plus an asynchronous LLM "overseer" polling every 30s) stays meaningful; the authors state plainly that "for recursively self-improving systems where the model weights are also updated, observability mechanisms, on their own, are clearly inadequate."

This is a concrete, measured instance of the fast, non-parametric loop described in Do self-improving agents really split into two distinct loops?: SICA's edits all land on scaffold components, and the paper supplies a safety rationale that survey's taxonomy leaves implicit — scaffold-only editing is "cheap and reversible" but also keeps the chain-of-thought legible in a way weight updates would not. Against the two dials in What separates self-improvement from policy improvement?, SICA sits on the self-referential side of the first dial — collapsing the ADAS meta/target split moves the improver fully inside the agent being improved — while the second dial stays anchored: the benchmark suite and utility function that define "best" are fixed and external to any agent in the archive. SICA is a worked example of GAI's "self-referential improver, externally anchored standard" cell, not a fully self-referential system by both dials at once.

The excerpt's Conclusion section is blank in the source text, and nothing here demonstrates that the 17%-to-53% jump compounds across further iterations rather than being a one-time capability gain from an already-sophisticated starting agent. The authors' own Limitations note that the agent struggled to generate "truely novel, innovative, feasible and interesting modification ideas," and that early low-quality ideas (their example: "fixating on caching open files") biased later suggestions toward variations on the same theme — a path-dependency that bounds how much compounding self-improvement this design can actually sustain, and which Does constraining edits make skill learning more stable? frames as exactly the failure mode that bounded, buffered editing is meant to prevent. The claim that scaffold-only editing is safer than weight-updating RSI is argued, not measured, in this excerpt; what is measured is a single-system capability gain under a fixed, human-set benchmark standard.