代理框架與評測
Recuris Freezes Agent Weights and Recursively Rewrites Skill Memory, Improving Results in 35 Long-Horizon Task Evaluations
Recuris separates task state, skill selection, and failure evidence into memory components that can be updated independently, then validates each modification against held-out data. The authors recorded improvements in 35 of 37 completed model–benchmark pairings, although some gains were not statistically significant and the evaluations also relied on LLM-simulated users and judges.
Recuris attempts to move agent “self-improvement” beyond model weights and a single monolithic prompt. It equips a frozen downstream agent with episodic memory and working memory: the latter preserves verified task state and selects skills based on the current gaps. At every step, the system records the state, retrieved experience, action, and observation, allowing failures to be attributed to specific memory components instead of producing only a final score.
The improvement loop uses a fixed Meta-Agent to inspect failure trajectories and modify only the relevant skill cards or routing components. Candidate modifications must pass a deterministic validation gate on paired held-out evidence before being written back to Skill Memory. The downstream model itself is not fine-tuned, nor does the model vote on whether its own modifications succeeded. The engineering value of this design is that memory packages can be versioned, rolled back, and reused across models, while avoiding indiscriminate retrieval over an ever-growing full chat history.
Across four settings—including τ²-Bench, SkillFlow, and Terminal-Bench 2.1—and ten models, the authors compared baseline agents with Recuris-equipped agents using `avg@4`, based on four trials per task. Of 37 completed pairings, 35 improved. On τ²-Retail, GPT-5.6 Sol gained 17.8 percentage points, while Claude Opus 5 gained 15.6 points to reach 87.9%. On SkillFlow, Qwen3.6-27B and Qwen3.6-35B improved by 16.6 and 13.5 points, respectively, with the gap reaching as much as 32.2 points in the longest interaction range.
These figures should not be interpreted as evidence that agents can already self-improve reliably and indefinitely. The confidence intervals for several small gains in the table include zero, while Granite declined by 0.3 points on SkillFlow. On Terminal-Bench, dynamic memory improved by only 2.3 points over fixed memory, which the authors acknowledge was not significant at the available sample size. τ² scoring also relies on a fixed LLM user simulator and an assertion judge, while reproducing some reported results requires model-family routing overrides selected using in-sample data. Future work should examine cross-organization workloads, memory contamination, and long-term cumulative costs, as well as whether the validation gate can prevent erroneous skills from spreading after multiple rounds of evolution.