AI News AI资讯 2h ago Updated 2h ago 更新于 2小时前 50

The Evolution of the Agent Harness 智能体框架的演进

AI agent capabilities saw a significant jump around Christmas 2025, caused not by model improvements alone but by the convergence of advancing model capabilities and maturing agent harnesses An "agent harness" encompasses everything beyond model weights—tools, context, memory, guardrails—that enables LLMs to interact with digital space rather than remaining isolated "brains in a vat" The evolution followed distinct phases: ReAct (prompting-based loops, Oct 2022), Toolformer (training tools in, F 2025年圣诞节前后AI Agent能力出现显著跃升,源于模型能力与Agent Harness(框架)改进曲线的交汇 Agent Harness是除模型权重外使Agent工作的所有组件,包括环境、工具、上下文、记忆和护栏 Agent发展经历了三个阶段:2022-2023年的"附加时代"(ReAct、AutoGPT)、2023-2024年的"人机协作退守"(Cursor/Copilot),以及2024年底后的能力逆转 核心观点:模型持续将Harness吸收进权重,工程师删除已被吸收的部分,最终剩余的是"为人注意力设计的Harness"而非为模型设计

68
Hot 热度
75
Quality 质量
72
Impact 影响力

Analysis 深度分析

TL;DR

  • AI agent capabilities saw a significant jump around Christmas 2025, caused not by model improvements alone but by the convergence of advancing model capabilities and maturing agent harnesses
  • An "agent harness" encompasses everything beyond model weights—tools, context, memory, guardrails—that enables LLMs to interact with digital space rather than remaining isolated "brains in a vat"
  • The evolution followed distinct phases: ReAct (prompting-based loops, Oct 2022), Toolformer (training tools in, Feb 2023), AutoGPT/BabyAGI (premature full autonomy, Spring 2023), and IDEs like Cursor/Copilot (retreat to human-in-the-loop, 2023-2024)
  • The critical gap between harness demands and model capabilities was widest during the AutoGPT era, where compounding errors over multi-step tasks produced failure rates as high as ~64% for 20-step tasks at 95% per-step reliability
  • The paradigm is shifting toward models absorbing harness logic into their weights, with engineers progressively deleting external harness components, leaving behind a harness designed for human attention management rather than model orchestration

Why It Matters

This framework reframes how AI practitioners should think about agent development—not as a model-only problem but as a systems problem where the interaction between model capability and harness design determines real-world performance. The historical analysis provides a roadmap for understanding why certain agent approaches succeeded or failed at different times, and the prediction that models will continue absorbing harness functionality into their weights has direct implications for how engineering teams should architect agent systems going forward.

Technical Details

  • Agent Harness Definition: The harness comprises the environment, tools, context management, memory/compaction systems, and guardrails/permissions that surround the model weights—essentially giving the model a "body" to perceive, act, persist information, and enforce boundaries in digital space
  • The Gap Model: Agent effectiveness is determined by the gap between two curves—what the harness demands of the model versus what the model can practically deliver. Closing this gap through synchronized improvement of both curves was the key to the Christmas 2025 breakthrough
  • ReAct (October 2022): Introduced the reasoning-acting-observing loop as a prompting technique external to model weights, establishing the conceptual foundation for agent loops before models were capable of executing them reliably
  • AutoGPT/BabyAGI Failure Mode (Spring 2023): The harness curve sprinted ahead of model capability, creating maximum gap width. Multi-step task reliability compounds multiplicatively—95% per-step reliability over 20 steps yields only ~36% task success—demonstrating why premature autonomy fails
  • Cursor/Copilot Strategy (2023-2024): Closed the gap by pulling harness demands below model capability, giving humans the agentic loop while the model accelerated human work. The o1 reasoning model (late 2024) marked the first inversion where model capability exceeded harness demands, signaling the beginning of the absorption phase

Industry Insight

  • Teams should resist the temptation to maximize autonomous agent loops before model capabilities have caught up; the Cursor/Copilot approach of human-in-the-loop orchestration remains the most reliable path to production value until models independently surpass harness complexity thresholds
  • The long-term trajectory suggests that significant portions of current harness engineering (tool calling, reasoning loops, memory management) will be absorbed into model weights, meaning engineering teams should invest in capabilities that are harder to bake into models—particularly human oversight interfaces, attention management, and strategic task decomposition
  • The "harness for human attention" thesis implies that the next competitive advantage in AI agent development will not come from more sophisticated external tooling but from better designs for how humans supervise, intervene in, and direct increasingly autonomous model behavior

TL;DR

  • 2025年圣诞节前后AI Agent能力出现显著跃升,源于模型能力与Agent Harness(框架)改进曲线的交汇
  • Agent Harness是除模型权重外使Agent工作的所有组件,包括环境、工具、上下文、记忆和护栏
  • Agent发展经历了三个阶段:2022-2023年的"附加时代"(ReAct、AutoGPT)、2023-2024年的"人机协作退守"(Cursor/Copilot),以及2024年底后的能力逆转
  • 核心观点:模型持续将Harness吸收进权重,工程师删除已被吸收的部分,最终剩余的是"为人注意力设计的Harness"而非为模型设计

为什么值得看

这篇文章为AI从业者提供了Agent技术演进的清晰历史框架,解释了为什么2025年底Agent能力突然质变,并预测了未来"模型吸收Harness"的长期趋势,对制定Agent开发策略具有重要参考价值。

技术解析

  • Agent Harness定义:指除模型权重外所有使Agent工作的组件,包括环境(perception)、工具(action)、记忆与压缩(persistence)、权限与护栏(boundaries),相当于给模型"赋予身体"
  • 能力曲线模型:用两条曲线描述Agent发展——Harness对模型的要求曲线 vs 模型实际能力曲线,两者之间的差距决定Agent有效性
  • 三阶段演进:①"附加时代"(2022-2023):ReAct提示技术、Toolformer训练工具调用概念、AutoGPT/BabyAGI过度自主导致差距最大;②"人机协作退守"(2023-2024):Cursor/Copilot将循环交给人类,Devin等全自主Agent成功率仅约15%;③"能力逆转"(2024年底):o1推理模型引入后模型能力首次超越Harness要求
  • 误差放大效应:循环不增加模型能力,而是放大现有能力;低于阈值时放大错误而非可靠性,如每步95%可靠性在20步任务中仅产生约36%成功率
  • 未来趋势:模型持续将Harness功能吸收进权重,工程师删除已被吸收的外部组件,最终Harness从"为模型设计"转变为"为人注意力设计"

行业启示

  • Agent开发策略调整:不应盲目追求全自主Agent,而应关注模型能力与Harness复杂度的匹配,2024年底后的能力逆转意味着现在可以重新评估自主Agent的可行性
  • 投资与研发重点:关注"模型吸收Harness"趋势,优先投资能提升模型内在能力的技术(如推理训练、工具调用内化),而非过度依赖外部框架复杂度
  • 产品定位建议:短期仍可借鉴Cursor/Copilot的人机协作模式,但中长期需为模型能力超越Harness要求的转折点做准备,重新设计Agent架构以利用这一趋势

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent LLM 大模型 Research 科学研究