CoEvo-RSI Aligns Models With Evolving Agent Harnesses

Researchers introduce CoEvo-RSI, a framework for recursive self-improving agents that coordinates changes to an agent’s execution harness and its underlying model. The paper addresses “brain–body mismatch”: a harness may require behaviors the model has not learned, or constrain capabilities it already has. CoEvo-RSI uses an archive of past designs and execution feedback so a meta-agent can develop harnesses, select or merge model checkpoints, and adapt models to the chosen harness.

The authors report that CoEvo-RSI uses only supervised fine-tuning on its own rejection-sampled trajectories. They say it makes Qwen3.5-9B competitive with larger models across multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following.

On TraceElephant, the framework surpassed DeepSeek-V4-Pro by 14.16 agent accuracy; on HomeBench VM, it surpassed GLM-5.3 by 4.90 success rate. The authors say these results support aligning model capabilities with changing execution requirements through bidirectional inheritance from historical designs. The paper was published on October 18 and submitted to the Hugging Face Papers page on October 21.

Researchers introduce CoEvo-RSI, a framework for recursive self-improving agents that coordinates changes to an agent’s execution harness and its underlying model. The paper addresses “brain–body mismatch”: a harness may require behaviors the model has not learned, or constrain capabilities it already has. CoEvo-RSI uses an archive of past designs and execution feedback so a meta-agent can develop harnesses, select or merge model checkpoints, and adapt models to the chosen harness.

The authors report that CoEvo-RSI uses only supervised fine-tuning on its own rejection-sampled trajectories. They say it makes Qwen3.5-9B competitive with larger models across multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following.

On TraceElephant, the framework surpassed DeepSeek-V4-Pro by 14.16 agent accuracy; on HomeBench VM, it surpassed GLM-5.3 by 4.90 success rate. The authors say these results support aligning model capabilities with changing execution requirements through bidirectional inheritance from historical designs. The paper was published on October 18 and submitted to the Hugging Face Papers page on October 21.