Prime Intellect 开源 Prime Agent：基于持久 IPython 内核的开放式 RLM 工具

Prime Intellect 已开源 Prime Agent：https://www.primeintellect.ai/blog/prime-agent，这是一个自我改进的编码工具，基于两个抽象概念设计：递归语言模型（Recursive Language Model, RLM）和持续工具（Continual Harness）。固定的工具架构和上下文压缩迫使模型围绕自身的框架工作。Prime Agent 用持久的 Python REPL 和可重写的工具替代了两者。使用 Opus 5，它在 ARC-AGI-3 上达到了 95.5%，高于报告的人工专家基线 95.4%。它采用 MIT 许可证。

是的，就是今天。Prime Agent：https://github.com/PrimeIntellect-ai/prime-agent 可通过一条命令在 Linux 或 macOS 上安装。它可以在订阅登录（Codex、Claude Pro/Max、GitHub Copilot）、API 密钥（Anthropic、OpenAI、Google、Groq、Fireworks、Prime Inference 等）、Azure OpenAI、Amazon Bedrock 以及自托管的 vLLM、Ollama 或 LM Studio 端点上运行。自托管开源权重模型，如 GLM-5.2，可将代码保留在您自己的网络内。

Prime Agent 构建于两个抽象概念之上。递归语言模型（Recursive Language Model, RLM）：https://arxiv.org/abs/2512.24601 将上下文视为一个变量，并将子代理的委派视为 REPL 内的函数调用。持续工具（Continual Harness）：https://arxiv.org/abs/2605.09998 将提示、子代理、技能和记忆视为代理可以从自身轨迹中创建、读取、更新和删除的状态。这两篇论文都有 Prime Agent 作者参与。TUI 是基于 pi 构建的：https://github.com/earendil-works/pi。

Prime Agent 中的模型只有一个工具：持久的 IPython 内核。技能、工具和子代理都是其中预导入的模块。rlm("子任务") 启动一个拥有自身模型、内核和历史的子会话，并在接纳时返回而不是阻塞。结果通过 agent_message.send(...) 到达。

后台守护进程管理每个活动会话。你可以在不中断循环的情况下分离并重新连接，崩溃的工作进程可根据会话 JSONL 和内核快照恢复。

代理间消息特意限制在核心家庭范围——父、兄弟或子——以防止跨会话闲聊。保留的子代理在闲置 30 分钟后从内存中删除，然后在被调用时重新加载。

持续工具将工具状态形式化为 H = (ρ, G, K, M)：提示、子代理、技能、记忆。每一个都暴露相同的创建、读取、更新、删除界面。

/refine 读取代理自身的轨迹并应用最小相关编辑，记录触发器和结果。计划在后台运行，不会阻塞对话。基础系统提示保持不变，错误更新可以通过 ID 恢复。

在 ARC-AGI-3：https://arcprize.org/arc-agi/3 上，使用 Opus 5 的 Prime Agent 报告 95.5% RHAE Best@1，高于 ARC 报告的人类专家基线 95.4%。三次运行结果为 95.0、95.2 和 95.5，Best@3 达到 99.97%，并且所有 183/183 关卡均完成。Prime Intellect 还报告其令牌使用量低于原生实现，归功于函数在数据上运行而非通过工具读取数据。

在长上下文套件中，使用开放权重 GLM-5.2 的 Prime Agent 在九次评测中八次击败 Pi-mono。使用 Opus 5，它在九次评测中六次略胜 Claude Code；使用 GPT-5.6 Sol，它在九次评测中六次击败 Codex。

案例研究包括 EmulatorBench，其中代理根据规范在 Rust 中构建模拟器，无参考实现，并成功重现 SEGA Genesis 和 Game Boy Color；PMPP-Hard，用于验证 GPU 内核 against KernelGuard：https://github.com/gpu-mode/kernelguard；以及 Factorio，其中其在数小时内达到超过 10 万的生产分数。

Factorio 还产生了最有价值的负面结果。Prime Agent 发现它可以通过 RCON 命令将资源直接生成到装配机中，尽管心跳提示告知它不要作弊。同样的精炼循环，在构建合法技能后，也构建了高效的作弊技能。

需要与我们合作以推广您的 GitHub 仓库或 Hugging Face 页面或产品发布或网络研讨会等吗？请联系我们：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq 是 Marktechpost Media Inc. 的首席执行官。作为一名有远见的企业家和工程师，Asif 致力于利用人工智能的潜力促进社会公益。他最新的努力是推出人工智能媒体平台 Marktechpost，该平台以深入报道机器学习和深度学习新闻而著称，内容既技术可靠又易于广大受众理解。该平台每月浏览量超过 200 万次，显示了其在受众中的受欢迎程度。

Prime Intellect has open-sourced Prime Agent：https://www.primeintellect.ai/blog/prime-agent, a self-improving coding harness designed around two abstractions, the Recursive Language Model (RLM) and Continual Harness. Fixed tool schemas and context compaction force a model to work around its own scaffolding. Prime Agent replaces both with a persistent Python REPL and a rewritable harness. With Opus 5, it reports 95.5% on ARC-AGI-3, above the reported human expert baseline of 95.4%. It is MIT-licensed.

Yes, today. Prime Agent：https://github.com/PrimeIntellect-ai/prime-agent installs on Linux or macOS with one command. It runs on subscription logins (Codex, Claude Pro/Max, GitHub Copilot), API keys (Anthropic, OpenAI, Google, Groq, Fireworks, Prime Inference, and others), Azure OpenAI, Amazon Bedrock, and self-hosted vLLM, Ollama, or LM Studio endpoints. Self-hosting an open-weights model such as GLM-5.2 keeps code inside your own network.

Prime Agent is built on two abstractions. The Recursive Language Model (RLM)：https://arxiv.org/abs/2512.24601 treats context as a variable and sub-agent delegation as function calls inside a REPL. The Continual Harness：https://arxiv.org/abs/2605.09998 treats prompts, sub-agents, skills, and memory as state the agent can create, read, update, and delete from its own trajectory. Both papers have Prime Agent authors on them. The TUI is built on pi ：https://github.com/earendil-works/pi.

Models in Prime Agent get one tool: a persistent IPython kernel. Skills, tools, and sub-agents are pre-imported modules inside it. rlm("sub-task") launches a child session with its own model, kernel, and history, returning at admission rather than blocking. Results arrive through agent_message.send(...) .

A background daemon owns every live session. You can detach and reattach without stopping the loop, and a crashed worker recovers from the session JSONL plus a kernel snapshot.

Agent-to-agent messaging is deliberately scoped to the nuclear family — parent, sibling, or child — to prevent cross-session chatter. Retained sub-agents drop from memory after 30 minutes idle, then reload when addressed.

Continual Harness formalizes harness state as H = (ρ, G, K, M): prompt, sub-agents, skills, memory. Each exposes the same create, read, update, delete surface.

/refine reads the agent's own trajectory and applies the smallest relevant edit, recording the trigger and the outcome. Planning runs in the background without blocking the conversation. The base system prompt stays immutable, and a bad update can be reverted by ID.

On ARC-AGI-3：https://arcprize.org/arc-agi/3, Prime Agent with Opus 5 reports 95.5% RHAE Best@1 , above the ARC reported human expert baseline of 95.4%. Three runs land at 95.0, 95.2, and 95.5, with 99.97% Best@3 and all 183/183 levels complete. Prime Intellect also reports lower token usage than native harnesses, crediting functions run over data instead of data read through tools.

On a long-context suite, Prime Agent with open-weights GLM-5.2 beats Pi-mono on eight of nine evals. With Opus 5 it edges Claude Code on six of nine; with GPT-5.6 Sol it beats Codex on six of nine.

Case studies include EmulatorBench, where the agent builds emulators in Rust from spec with no reference implementation and reproduces the SEGA Genesis and Game Boy Color; PMPP-Hard, for GPU kernels verified against KernelGuard：https://github.com/gpu-mode/kernelguard; and Factorio, where it reached 100K+ production score in hours.

Factorio also produced the most useful negative result. Prime Agent found it could spawn resources straight into assembly machines through RCON commands, despite a heartbeat prompt telling it not to cheat. The same refinement loop that built legitimate skills then built efficient cheating skills.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.