Prime Agent: A Self-Improving RLM Harness

ID: 25fd4d45
原文链接: https://arxiv.org/abs/2608.23552
PDF: https://arxiv.org/pdf/2608.23552v1
作者: S, e, t, h, , K, a, r, t, e, n, ,, , A, l, e, x, , L, ., , Z, h, a, n, g, ,, , K, e, v, i, n, , T, h, o, m, a, s, ,, , S, e, b, a, s, t, i, a, n, , M, ü, l, l, e, r, ,, , E, l, i, e, , B, a, k, o, u, c, h, ,, , D, a, n, i, e, l, , A, u, r, a, s, ,, , M, i, k, a, , S, e, n, g, h, a, a, s, ,, , F, a, r, e, s, , O, b, e, i, d, ,, , K, o, n, s, t, a, n, t, i, n, , D, u, n, a, s, ,, , J, o, h, a, n, n, e, s, , H, a, g, e, m, a, n, n, ,, , S, a, m, i, , J, a, g, h, o, u, a, r
日期: 2026-08-24
更新: 2026-08-24
分类: agents
来源类型: paper
标签: prime-agent, rlm, agent-harness, arc-agi, test-time-compute
质量评分: 5/5
抓取时间: 2026-08-26T04:27:05Z

中文导读

Prime Agent 是开源的长视野评测与 coding-agent harness：持久 IPython REPL 承担 Recursive Language Model 抽象，负责程序化的上下文处理与 test-time compute；Continual Harness 让历史记忆技能prompt子 agent 规格跨轨迹保留；递归子 agent 之间直接通信，Agents View 供人检查和接管 daemon 化的会话harness 统一处理执行恢复验证与资源记账，把策略构造留给模型本身，避免 harness 失败被记成模型失败效果上，ARC-AGI-3 RHAE Best@1 从 30% 提到 95.5%，在长上下文编码GPU kernel 生成模拟器构建自主 nanoGPT speedrun 等任务上追平或超过原生与主流 harness；Factorio 上精炼循环支撑连续科技演进专职子 agent 支持并行分工代码开源于 github.com/PrimeIntellect-ai/prime-agent

为什么值得关注

开源 RLM harness：持久 IPython REPL + 跨轨迹记忆/技能，把 ARC-AGI-3 RHAE Best@1 从 30% 拉到 95.5%

对做 agent harness 与长视野评测的工程师，这是一个可直接上手的开源参考实现：持久 IPython REPL、跨轨迹保留的记忆与技能、人可检查接管的 Agents View，代码已在 GitHub 开源（PrimeIntellect-ai/prime-agent）。

关键信息

论文标题: Prime Agent: A Self-Improving RLM Harness
作者: S, e, t, h, , K, a, r, t, e, n, ,, , A, l, e, x, , L, ., , Z, h, a, n, g, ,, , K, e, v, i, n, , T, h, o, m, a, s, ,, , S, e, b, a, s, t, i, a, n, , M, ü, l, l, e, r, ,, , E, l, i, e, , B, a, k, o, u, c, h, ,, , D, a, n, i, e, l, , A, u, r, a, s, ,, , M, i, k, a, , S, e, n, g, h, a, a, s, ,, , F, a, r, e, s, , O, b, e, i, d, ,, , K, o, n, s, t, a, n, t, i, n, , D, u, n, a, s, ,, , J, o, h, a, n, n, e, s, , H, a, g, e, m, a, n, n, ,, , S, a, m, i, , J, a, g, h, o, u, a, r
arXiv: https://arxiv.org/abs/2608.23552
发布时间: 2026-08-24
arXiv 分类: c, s, ., A, I, ,, , c, s, ., C, L, ,, , c, s, ., S, E
备注: 16 pages, 10 figures. Technical report. Code: https://github.com/PrimeIntellect-ai/prime-agent
关联标签: prime-agent, rlm, agent-harness, arc-agi, test-time-compute

English Abstract

Language models are sequential processors, but long-horizon agency requires external information and computation beyond model weights and active context. Prime Agent is an open-source harness for long-horizon evaluation and coding-agent workflows. A persistent IPython REPL follows the Recursive Language Model abstraction for programmatic context processing and test-time compute, while Continual Harness preserves histories, memories, skills, prompts, and subagent specifications across trajectories. Recursive subagents coordinate through direct agent-to-agent communication, and the Agents View lets humans inspect and manage daemon-backed sessions. Prime Agent standardizes execution, recovery, verification, and resource accounting while leaving strategy construction to the model. This low-friction, expressive membrane prevents harness failures from becoming model failures and pushes measurement toward the model's true maximal underlying capability. Prime Agent raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5% and matches or exceeds native and popular harnesses across long-context coding, GPU-kernel generation, emulator construction, and autonomous nanoGPT speedruns. On Factorio, we find refinement allows for continuous technology progression and dedicated subagents enable parallelized work. Code is available at https://github.com/PrimeIntellect-ai/prime-agent.

English Summary

Prime Agent is an open-source harness for long-horizon evaluation and coding-agent workflows. A persistent IPython REPL follows the Recursive Language Model abstraction for programmatic context processing and test-time compute; Continual Harness preserves histories, memories, skills, prompts, and subagent specifications across trajectories; recursive subagents coordinate through direct agent-to-agent communication; an Agents View lets humans inspect and manage daemon-backed sessions. The harness standardizes execution, recovery, verification, and resource accounting while leaving strategy construction to the model, so harness failures do not masquerade as model failures....

Obsidian Notes

内容由 opencli arxiv paper 拉取 arXiv 元数据与摘要生成。
中文导读与价值判断锚定在条目已有摘要、论文摘要、作者、日期与分类信息上；未补充论文摘要之外的实验细节。
本页由 AAIF content-fetcher 定时任务生成（2026-08-26），仅新增内容页，未改动 entries.json。