RSIHub Packs 4 Agent Evolution Methods Into One Framework: Up to +13.5 pp

Research on self-improving agents keeps running into the same chore: every team rewrites the same scaffolding for runs, evaluation and version tracking. Zimu Wang's team has introduced RSIHub, an open-source framework that moves this scaffolding into shared infrastructure. The authors ran four agent evolution methods on it and compared them on two benchmarks. The largest gain in overall pass@1 was 13.5 percentage points. The main takeaway is more modest than the headline suggests: there is no one-size-fits-all winner.

Agent evolution as a set of swappable parts

Agent evolution works like this: the agent performs tasks, the system collects feedback, and that feedback is used to rewrite prompts, skills, tools or the orchestration scheme. The idea is simple, but implementations are usually built for one specific paper and are incompatible with each other.

In RSIHub, a method is described as a "recipe." A recipe is assembled from reusable operators and explicitly declares which resources may be changed. The platform handles everything else: execution, evaluation, failure recovery and traceability of candidate origins (lineage). The researcher writes only the evolution policy, that is, decides what to mutate and which candidates to keep.

The key architectural decision concerns evaluation. It sits outside the mutable surface, so the agent cannot rewrite its own examiner. For systems that edit themselves, this guards against the most obvious form of cheating. Each result is tied to a specific candidate and its evaluation conditions, so the history of any improvement can be traced back.

Four methods on one substrate

The authors implemented the methods without a single change to the framework core. All method-specific logic stayed in recipes and operators. This is the main practical test of composability.

• AHE — the first agent evolution method they reproduced.
• Hyperagents — an approach to self-modifying agents.
• A-Evolve — an agent evolution method based on feedback.
• GEPA — an optimization method based on reflection over execution results.

The evaluation setup looked like this:

• Agent harnesses: MiniSWE and Codex.
• Benchmarks: Terminal Bench 2 (terminal tasks) and τ³ Banking (banking scenarios).
• Conditions: fixed model and evaluator configurations, so that the difference is determined only by the evolution method.
• Main metric: overall pass@1.
• Best result: +13.5 pp pass@1.

Where the numbers contradict expectations

The most interesting part of the results is not the record but its limits. The ranking of methods changes from setup to setup. A method that leads with one harness or benchmark can lose on another. Gains on the tasks used for optimization do not always hold on held-out tasks. This is classic overfitting, except that it now happens at the level of prompts, skills and tools rather than weights.

For the industry, this is an unwelcome but useful signal. Papers on agent self-improvement often compare their own method against someone else's on their own scaffolding and with their own assumptions. A shared protocol makes such comparisons simpler. It also exposes cases where a gain depends on a lucky match of conditions.

What to do with this

Teams building coding agents or enterprise assistants can use RSIHub as a testbed. The recipes for the four methods are already available, and teams can run them on their own tasks instead of writing the scaffolding from scratch.

Practical rules follow from the results themselves. Keep a held-out set and evaluate on it, not only on the optimization tasks. Do not rely on the ranking from a single benchmark. Store the lineage of every candidate to understand which change produced the gain. The code is open on GitHub (simple-agent-lab/RSIHub), the project has already gathered 186 stars, and the authors are asking for contributions in the form of new evolution methods and benchmarks. If new recipes appear in the ecosystem, the community will have a common way to check which of them work beyond a single paper.