Build self-improving agent system with Fable 5 in 14 steps loops, dynamic workflows, routines 用Fable 5在14个步骤中构建自我改进的代理系统：循环、动态工作流、例程

Most people are using Claude Fable 5 like Sonnet 4.6 with a bigger context window. They prompt it. It works for 5 minutes. They close the tab. 大多数人使用 Claude Fable 5 时，就像把它当成一个上下文窗口更大的 Sonnet 4.6 来用。他们给它提示词，它能正常运行 5 分钟，然后他们就关掉标签页了。

9 out of 10 users have never run an agent system that compounds - where every run leaves the next run smarter, every state file accumulates, every skill sharpens.每10个用户中就有9个从未运行过能够实现复合增长的智能体系统——在这种系统中，每一次运行都会让下一次运行变得更加智能，每一个状态文件都在积累，每一项技能都在精进。

Fable 5 was built to run for days. You’re using it for minutes. This is the 14-step roadmap to build the self-improving system Fable 5 was designed for.Fable 5 就是为连续运行数天而设计的。而你目前只用了它几分钟。这份14步路线图将为你构建出 Fable 5 所设计的自我改进系统。

Follow my Substack to get fresh AI alpha: 关注我的 Substack，获取最新 AI 阿尔法资讯：movez.substack.commovez.substack.com

Claude Fable 5 launched June 9, 2026 - the first publicly available Mythos-class model, the tier Anthropic put one rung above Opus. Claude Fable 5 于2026年6月9日发布——这是首个向公众开放的Mythos级模型，也是Anthropic将其定位在Opus之上的一个层级。

This is the 14-step roadmap to build the self-improving system Fable 5 was designed for - sourced from Anthropic engineering posts, the team’s public experiments, and verified against the launch documentation as of June 2026. 这是构建自改进系统Fable 5的14步路线图——资料来源包括Anthropic的工程博客、团队公开实验，并已对照2026年6月的发布文档进行验证。

Three tiers: what Fable 5 actually unlocks, the three primitives that make it compound (loops, dynamic workflows, routines), and the self-improvement layer that turns it into a system. 三个层级：Fable 5真正解锁的能力、使其产生复合效应的三大原语（循环、动态工作流、例行程序），以及将这一切转化为系统的自我改进层。

14 steps. 3 tiers. Stop prompting. Start building a system that compounds. 14个步骤。3个层级。停止提示。开始构建能够持续迭代的系统。

PART 1 · What Fable 5 actually unlocks 第一部分 · Fable 5真正解锁的能力

01. Fable 5 is a Mythos-class model. Days-long autonomy is the headline.01. Fable 5 属于 Mythos 级别模型。其最大亮点是长达数日的自主运行能力。

Claude Fable 5 launched June 9, 2026 as the first publicly available Mythos-class model - the tier Anthropic introduced one rung above Opus. Claude Fable 5 于 2026 年 6 月 9 日发布，是首款公开可用的 Mythos 级别模型——这是 Anthropic 在 Opus 之上新增的更高层级。

Mythos Preview shipped in April through Project Glasswing to a handful of critical-infrastructure partners; Fable 5 is the version Anthropic considered safe for general release, with built-in safety classifiers that decline requests in high-risk areas. Mythos 预览版于 4 月通过 Project Glasswing 向少数关键基础设施合作伙伴开放；而 Fable 5 是 Anthropic 认为可安全公开发布的版本，内置了可在高风险领域拒绝请求的安全分类器。

Mythos 5 (without those classifiers) remains Glasswing-only. Mythos 5（未包含上述分类器）仍仅限 Project Glasswing 使用。

What Fable 5 actually does that previous Claude models couldn’t sustain, from Anthropic’s launch documentation: 从Anthropic的发布文档来看，Fable 5真正实现了此前Claude模型无法持续做到的功能：

• Days-long autonomous sessions. Run inside an agent harness like Claude Code or Claude Managed Agents (CMA), Fable 5 can work for days - planning across stages, delegating to sub-agents, and checking its own work.长达数天的自主运行会话。在Claude Code或Claude托管代理（CMA）等代理框架中运行时，Fable 5可以持续工作数天——分阶段规划任务、将工作委派给子代理、并自主核查自身工作成果。
• Self-verification built in. Writes its own tests to check its work. Uses vision to check outputs against goals. Distills lessons into general rules. Tests its own assumptions.内置自我验证机制。自主编写测试脚本来检验工作成果，利用视觉能力将输出结果与目标进行比对，将经验教训提炼为通用规则，并测试自身假设的合理性。
• Most ambitious code work. Large migrations, complex implementations, multi-day autonomous coding sessions. The headline use case Anthropic puts forward is “hand off large projects and review completed deliverables.”最具雄心的代码工作。大规模迁移、复杂功能实现、持续数天的自主编程会话。Anthropic主推的典型应用场景是”承接大型项目并交付经审核的完整成果”。
• Multi-stage knowledge work. Deep research and analysis to deliverables ready for review - with minimal oversight.多阶段知识工作。从深度调研分析到可交付成果供审阅——仅需极少监督。

The pricing matches the tier: $10 per million input tokens, $50 per million output tokens, with the existing 90% input token discount for prompt caching. 定价与层级匹配：输入每百万Token 10美元，输出每百万Token 50美元，Prompt缓存可享受现有90%输入Token折扣。

Available on Claude API, AWS, Amazon Bedrock, Vertex AI, Microsoft Foundry, and the consumption-based Enterprise plan. This is not a subscription model. Heavy use earns its own bill.适用于Claude API、AWS、Amazon Bedrock、Vertex AI、Microsoft Foundry及基于消耗量的企业计划。此非订阅模式。高强度使用将产生独立账单。

02. Self-improving is not self-learning. 02. 自我改进并非自我学习。

The phrase “self-improving agent system” gets thrown around carelessly. The version that’s real and the version that’s hype are very different things, and the gap is worth understanding before you build anything. “自我改进型代理系统”这个词常被轻率地提及。现实版本与炒作版本截然不同，在构建任何系统之前，理解这一差距至关重要。

• Self-learning - the agent updates its own weights based on what it learns. Fable 5 does not do this. No publicly available model does this in production. Recursive self-improvement (RSI) is the long-term direction Anthropic itself warned about in May 2026, not the capability shipping today.自我学习——代理会根据所学内容更新自身权重。Fable 5 并未实现此功能。目前没有任何公开可用的模型在生产环境中做到这一点。 递归自我改进（RSI）是 Anthropic 自身在 2026 年 5 月警告过的长期方向，而非当前已具备的能力。
• Self-improving - the system around the agent compounds. Each session writes lessons to memory. Skills sharpen as edge cases get added. State files accumulate verified facts. Eval loops refine prompts and rubrics. The model stays the same; the environment it runs in gets sharper.自我改进——围绕代理的系统在持续积累。每次会话都会将经验写入记忆。随着边缘情况的增多，技能也变得更加精准。 状态文件积累已验证的事实。评估循环优化提示词和评分标准。模型本身保持不变，但其运行环境变得更加敏锐。

Self-improvement, in this sense, is a property of the system you build. Fable 5 has the raw capability - long context, sub-agent delegation, vision self-check, days-long stamina - that turns the environment-feedback loop into something that actually compounds run over run. 在这个意义上，自我改进是你所构建系统的一个属性。Fable 5 拥有原始能力——长上下文、子代理委派、视觉自我检查、持续数天的耐力——能将环境反馈循环转化为每次运行中真正实现累积增长的东西。

Anthropic’s engineering team puts it directly: Anthropic 的工程团队直接指出：

“Rather than directly prompting and steering Fable 5, it’s often better to design loops that let the model self-correct in response to environment feedback (e.g., /goal or Outcomes) and manage its own context (e.g., via memory).” “与其直接提示和引导 Fable 5，通常更好的做法是设计循环，让模型能够根据环境反馈（例如 /goal 或 Outcomes）自我纠正，并管理自己的上下文（例如通过记忆）。”

03. The compound stack: four layers, one feedback loop. 03、复合堆栈：四层结构，一个反馈循环。

Figure 1 at the top of this article shows the architecture in one diagram. Read it from the bottom up - that’s the order the system gets built, and the order the leverage compounds. 本文顶部的图 1 以一张图展示了该架构。请从下往上阅读——这是系统构建的顺序，也是杠杆效应逐层叠加的顺序。

• Layer 1 · Primitives. Fable 5 itself, sub-agents, worktrees, the tools the agent reaches for. Raw capability with no system around it yet. This is what most people use today.第一层 · 基础组件。Fable 5 本身、子智能体、工作树、智能体可调用的工具。原始功能，尚无系统化编排。这是目前大多数人使用的层面。
• Layer 2 · Orchestration. /goal and Outcomes for self-correcting loops. Dynamic Workflows for complex multi-step orchestration. Routines for laptop-off cloud runs. This is what turns the primitives into a workflow.第二层 · 编排。/goal 与 Outcomes 用于实现自我修正循环。动态工作流用于复杂多步骤编排。日常程序用于离线云端运行。这一层将基础组件转化为实际工作流。
• Layer 3 · Memory. State files, Skills, Knowledge Bases, lessons written down. Memory is what makes tomorrow’s session resume instead of restart.第三层 · 记忆。状态文件、技能库、知识库、书面经验记录。记忆能让明天的会话从暂停处继续，而非重新开始。
• Layer 4 · Self-improvement. Vision self-checks, eval loops, rule distillation. The agent grades its own output, refines the Skill that produced it, writes the lesson back to memory. The loop closes.第四层 · 自我改进。视觉自检、评估循环、规则提炼。智能体对自己的输出进行评分，优化产生该输出的技能，将经验教训写回记忆。循环闭环。

The reason this architecture compounds: every output from layer 1 flows up through layer 4, where it gets graded, distilled, and written back to layer 3. Tomorrow’s run at layer 1 inherits the sharpened memory and refined Skills from yesterday. The model is stateless; the system around it isn’t.这种架构之所以能产生复利效应：第一层的每个输出都会向上流向第四层，在那里被打分、提炼，并写回第三层。第二天第一层的运行继承了前一天优化后的记忆和精炼的技能。模型本身无状态，但其周围的系统并非如此。

04. When to use Fable 5 vs Opus 4.8 vs Sonnet 4.6. The cost-capability matrix.04. 何时使用 Fable 5、Opus 4.8 与 Sonnet 4.6。成本-能力矩阵。

Fable 5 costs ~5× what Opus 4.8 does per token. Not every step in a self-improving system needs the top tier. The teams running this in production route by task complexity, not by default: Fable 5 每 token 的成本大约是 Opus 4.8 的 5 倍。在自我改进系统中，并非每个步骤都需要最顶级的模型。生产环境下运行的团队按任务复杂度进行路由，而非默认使用最高配置：

• Fable 5 for the heavy-lift orchestrator role: planning across days, delegating to sub-agents, checking work with vision, distilling rules from accumulated evidence. Use Fable 5 where the “days at a time” capability earns its pricing. Fable 5 扮演繁重协调者角色：进行跨数天规划、向子代理分派任务、通过视觉功能检查工作、从积累证据中提炼规则。当“数日连续运行”能力值得其定价时，选用 Fable 5。
• Opus 4.8 for hard-but-bounded subtasks the orchestrator delegates: architecture decisions, complex debugging, deep code reviews. Also the explicit fallback for any request Fable 5’s classifiers block (cyber, bio, chem, distillation). Opus 4.8 负责协调者分派的困难但有边界子任务：架构决策、复杂调试、深度代码审查。同时作为 Fable 5 分类器拦截请求（涉及网络、生物、化学、蒸馏领域）的显式后备方案。
• Sonnet 4.6 for high-volume worker tasks: lint passes, simple refactors, test scaffolding, doc updates. The bulk of fan-out work runs here. Sonnet 4.6 承担高容量执行任务：代码格式检查、简单重构、测试框架搭建、文档更新。大部分工作负载扩散任务在此处运行。
• Haiku 4.5 for grader sub-agents and cheap classifiers. Independent context window, low cost - ideal for the verifier role Anthropic explicitly recommends.Haiku 4.5 用于评分子代理和低成本分类器。独立上下文窗口、低开销——完美适配 Anthropic 明确推荐的验证者角色。

The cost pattern that makes a self-improving system economical, used by teams running this in production: orchestrator on Fable 5, workers on Sonnet 4.6, graders on Haiku 4.5, fallback to Opus 4.8 on classifier blocks. Same pattern Anthropic engineers use internally.使自我改进系统经济实惠的成本模式，已被生产环境中的团队采用：使用Fable 5作为编排器，Sonnet 4.6执行工作，Haiku 4.5进行评估，在分类器模块中回退到Opus 4.8。这与Anthropic工程师内部使用的模式相同。

PART 2 · The three Primitives 第二部分 · 三大原语

05. /goal vs Outcomes. Two implementations of the same idea.05. /目标与成果。同一思想的两种实现方式。

The Anthropic Claude Code team publishes two near-identical primitives for goal-driven loops one in each harness. Anthropic的Claude Code团队发布了两种几乎相同的目标驱动循环原语——每种框架各一种实现。

They share the same shape: an independent grader checks the work, a not-met verdict starts the next iteration, the loop exits when the grader passes. 它们共享相同的结构：由独立的评分器检查作品，若结果为“未通过”则开始下一次迭代，当评分器判定通过时循环结束。

The implementations differ in surface details that matter for which you use. 不同实现方式在表面细节上存在差异，而这些细节决定了你应选用哪种方案。

The decision rule between them is short: 两者之间的决策规则很简短：

• Use /goal in Claude Code when the work happens at your machine and you want a quick, in-session loop with a measurable end state. Best for hands-on coding, debugging flaky tests, refining a single file. Plain text goal, model grader, in-terminal feedback.当工作在本地机器上进行，且希望快速的会话内循环并具备可衡量的最终状态时，请在 Claude Code 中使用 /goal 指令。最适合动手编码、调试不稳定的测试、优化单个文件。使用纯文本目标、模型评分器、终端内反馈。
• Use Outcomes in CMA when the work needs to run for hours or days on Anthropic-hosted infrastructure with a sandbox, GPUs, or a controlled environment. Best for ML training, long-running migrations, multi-day research. File-based rubric with gradable criteria, sub-agent grader, hard max_iterations bound.在Anthropic托管的基础设施（配有沙箱、GPU或受控环境）上执行需要运行数小时或数天的工作时，请在CMA中使用Outcomes方案。该方案最适合机器学习训练、长时间运行的迁移任务以及跨数天的研究项目。采用基于文件的评估标准（包含可评分的指标、子代理评分器以及硬性最大迭代次数限制）。

Both share the structural move that makes them work: the agent that wrote the code is not the agent that grades it. We go deeper on why that matters in step 6.两者都依赖相同的结构性调整来确保有效性：编写代码的代理与评分的代理并非同一实体。我们将在第6步深入探讨这一区别的重要性。

06. Verifier sub-agent beats self-critique. 06. 验证子代理优于自我批评。

Anthropic engineer Prithvi Rajasekaran wrote a piece on the engineering blog showing models have a hard time self-critiquing their own outputs. The Claude Code team confirmed this empirically with Fable 5: Anthropic工程师Prithvi Rajasekaran在工程博客上发表了一篇文章，指出模型在自我批评自身输出方面存在困难。Claude Code团队通过Fable 5实证验证了这一结论。

“We’ve found that a verifier sub-agent tends to outperform self-critique with Fable 5” “我们发现，在Fable 5中，使用验证子代理通常比自我批判效果更好”

The mechanism is structural, not about “trying harder.” A model evaluating its own output sees its own reasoning trail and prefers conclusions consistent with what it already wrote. 其机制本质上是结构性的，而非关乎“更加努力”。模型评估自身输出时，会看到自己的推理轨迹，并倾向于选择与已写内容一致的结论。

A separate model evaluating the same output sees only the artifact and the rubric. **The verifier has no skin in the maker’s game.**而另一个模型评估同一输出时，看到的只是成品和评分标准。验证者与创作者的“游戏”毫无利益关联。

What the chart actually shows, beyond the headline numbers: 图表所展示的，除了醒目的数据之外，还有更深的含义：

• Fable 5 made larger structural changes - TRAIN_SEQ_LEN=2048 train+eval (−0.0179), overlapped sliding-window eval (−0.0207), int6 QAT + int6 expo (−0.0163). Each is an architecture-level move, not a constant tweak.Fable 5 做出了更大的结构性调整——TRAIN_SEQ_LEN=2048 训练+评估（−0.0179）、重叠滑动窗口评估（−0.0207）、int6 量化感知训练 + int6 指数（−0.0163）。每一步都是架构层面的操作，而非常数微调。
• Fable 5 pushed through a quantization regression to its biggest win - instead of reverting after a failed experiment, it continued investigating.Fable 5 在量化回归中冲向了最大胜利——它没有在一次失败实验后回退，而是继续深入探究。
• Opus 4.7’s first experiment (QK_GAIN_INIT=5.0) produced a small win. Nearly everything that followed used the same template: adjust a scalar, measure, keep if positive. The shape is safer, not better.Opus 4.7 的第一次实验（QK_GAIN_INIT=5.0）带来了一小步胜利。此后几乎所有实验都沿用同一模板：调整标量、测量、保留正值。这种模式更稳妥，却非更优。

The takeaway for system design: Fable 5 with an independent verifier explores larger hypothesis spaces and recovers from negative intermediate results. Without the verifier, the same model has nothing forcing it past the first “good enough.”对系统设计的启示：配备独立验证器的 Fable 5 能够探索更大的假设空间，并从负面的中间结果中恢复。而没有验证器时，同样的模型没有任何力量迫使它突破第一个“足够好”的状态。

07. Dynamic Workflows compose self-correction patterns. 07. 动态工作流构成自我修正模式。

Dynamic Workflows shipped in Claude Code on May 28, 2026. 动态工作流于2026年5月28日在Claude Code中发布。

The idea: Claude writes its own JavaScript harness on the fly - a file with agent(), parallel(), and pipeline() primitives, plus standard JS to process the data flowing between them. The harness is custom-built for the task, not generic. 核心理念：Claude实时编写自己的JavaScript驱动框架——一个包含agent()、parallel()和pipeline()原语，以及标准JavaScript来处理它们之间流动数据的文件。该驱动框架是为特定任务量身定制的，而非通用型。

5月29日

Excited to share our most powerful new Claude Code feature: dynamic workflows! Mention “workflow” in a prompt and Claude will dynamically create an orchestration plan that it strictly follows, allowing you to confidently trust that every stage happens in the right order even

For self-improving systems with Fable 5, three of the six documented Dynamic Workflow patterns earn their place: 在Fable 5的自我改进系统中，六个已记录的动态工作流模式中有三个占据了重要地位：

• Fan-out-and-synthesize. Split the work into N independent pieces, run an agent on each in parallel, synthesize results. Best when each step benefits from its own clean context window - e.g., evaluating each rule in a Skill against historical examples.扇出与综合。将工作拆分为N个独立片段，并行运行各自Agent处理，最后综合结果。最适合每个步骤需要独立上下文窗口的场景——例如，对照历史案例逐个评估技能中的每条规则。
• Adversarial verification. For each maker agent, spawn an independent verifier with no exposure to the maker’s reasoning. The structural fix for self-preferential bias from step 6, applied per task.对抗式验证。为每个生产Agent并行生成一个独立的验证器，且该验证器不接触生产者的推理过程。这是针对第6步自我偏好偏差的结构化修正方案，按任务粒度应用。
• Loop until done. Loop spawning agents until a stop condition is met -no new findings, no more errors in the logs, theory verified. Pair with /goal to set a hard completion requirement.循环直至完成。持续生成Agent直到满足终止条件——无新发现、日志无错误、理论验证通过。配合/goal设定硬性完成要求。

The two patterns that don’t typically appear in self-improving systems but are worth knowing: classify-and-act (route the task to the right model based on a classifier) and tournament (pairwise comparison for taste-based ranking). The first is useful for model routing (step 4). The second is rare in coding loops but useful for design or naming tasks.两种在自优化系统中不常见但值得了解的范式：分类执行（通过分类器将任务路由至正确模型）与锦标赛制（两两对比实现审美排序）。前者适用于模型路由（第4步）。 后者在编程循环中罕见，但对设计或命名任务效果显著。

08. Worktrees for parallel safety. Days-long Fable 5 sessions, no file collisions.08. 工作树实现并行安全。长达数天的 Fable 5 会话，无文件冲突。

The moment a self-improving system spawns more than one agent, files start colliding. Two agents writing the same file is the same problem as two engineers committing to the same lines without talking first. 当自我改进系统产生多个智能体时，文件便开始发生冲突。两个智能体同时写入同一个文件，与两个工程师未经事先沟通便提交同一段代码是同样的问题。

A git worktree fixes it - a separate working directory on its own branch sharing the same repo history, so one agent’s edits literally cannot touch the other’s checkout. Git 工作树可以解决这一问题——它提供一个共享同一仓库历史但位于独立分支上的单独工作目录，因此一个智能体的编辑内容实际上无法触及另一个智能体的检出目录。

For self-improving systems where Fable 5 spawns sub-agents to verify or specialize, worktrees are non-optional: 对于由 Fable 5 生成子智能体进行验证或专业化的自我改进系统而言，工作树是不可或缺的：

• Maker writes in worktree A. Verifier reads in worktree B (or runs against the worktree A checkout with read-only filesystem). No risk the verifier’s exploration touches the maker’s state.制作者在工作树A中编写。验证者在工作树B中读取（或针对工作树A的检出运行只读文件系统）。验证者的探索不会触及制作者的状态。
• Parallel structural experiments. If Fable 5 explores multiple architecture changes (like in Parameter Golf), each experiment runs in its own worktree. The orchestrator collects results from all of them; the best one merges.并行结构实验。如果Fable 5探索多个架构变更（如在参数高尔夫中），每个实验在自己的工作树中运行。编排者收集所有实验的结果；最佳结果会被合并。
• Days-long runs with checkpoints. Each major phase can be a separate worktree. A failed phase doesn’t poison the rest.带有检查点的数日运行。每个主要阶段可以拥有独立的工作树。失败的阶段不会污染其他部分。

In Claude Code, worktrees are exposed three ways: git worktree directly, a —worktree flag to open a session in its own checkout, and an isolation: worktree setting on subagents so each helper gets a fresh checkout that cleans itself up after the session ends. 在Claude Code中，工作树通过三种方式暴露：直接使用git worktree，使用—worktree标志在独立检出中打开会话，以及为子代理设置isolation: worktree选项，使每个辅助代理获得一个在会话结束后自动清理的新检出。

09. Routines for days-long orchestration. Laptop closed. Fable 5 working.09. 可执行数日工作流的例程。笔记本电脑合上，Fable 5 仍在运行。

Routines launched April 14, 2026 in research preview. They’re saved Claude Code configurations - a prompt, repositories, connectors, permissions - that run on Anthropic-managed cloud infrastructure on a trigger. 这些例程于 2026 年 4 月 14 日以研究预览形式上线。它们是保存好的 Claude 代码配置——包含提示词、代码库、连接器和权限——在 Anthropic 管理的云基础设施上按触发条件执行。

Your laptop can be off. The run still happens.你的笔记本电脑可以关机。运行过程仍会照常进行。

For Fable 5 specifically, Routines are the trigger layer that earns the model’s capability. Anthropic measures Fable 5’s “days at a time” on Claude Managed Agents - a hosted sandbox with full tools and no local machine constraint. 对于 Fable 5 而言，这些例程是激发模型能力的触发层。Anthropic 在 Claude 托管代理（一个配备完整工具集、无本地机器限制的托管沙盒环境）上，以“持续运行数日”为基准评估 Fable 5 的表现。

The Parameter Golf experiment ran for up to 8 hours on 8×H100 GPUs. That class of run doesn’t happen on your laptop. 参数高尔夫实验在8×H100 GPU上持续运行了长达8小时。这类规模的实验不是你的笔记本能跑得动的。

The three Routine trigger types, mapped to self-improvement patterns: 三种Routine触发类型，对应着自我改进模式：

• Schedule triggers - the morning briefing pattern. Daily at 7am: re-run yesterday’s eval suite, distill any new failure modes into Skills, write the digest to Slack. The agent gets sharper while you sleep.计划触发 - 晨间简报模式。每天早上7点：重新运行昨日的评估套件，将新出现的故障模式提炼为技能，将总结写入Slack。在你睡眠时，智能体悄然变得更敏锐。
• API triggers - the “fire on event” pattern. CI fails → fire a Routine to investigate. Sentry alert → fire a Routine to triage. The self-improving system reacts to your real environment, not a fixed schedule.API触发 - “事件驱动”模式。CI失败 → 触发Routine进行调查。Sentry告警 → 触发Routine进行分类。这个自我改进系统响应的是你的真实运行环境，而非固定时间表。
• GitHub event triggers - the “learn from real work” pattern. On PR open, run an evaluation against the latest Skills. On merge, write any new patterns the PR introduced back to the Skill. Repository state and Skill state stay in sync.GitHub 事件触发——“从实际工作中学习”模式。当 Pull Request 开启时，针对最新技能运行评估。当合并时，将 PR 引入的任何新模式写回技能库。仓库状态与技能库状态保持同步。

> /schedule daily at 7am, use Fable 5 in CMA
  Goal: Re-run yesterday’s eval suite against the latest skills.
  Any test that newly passes → distill the pattern into the skill.
  Any test that newly fails → investigate, document in STATE.md.
  Post the digest to #engineering. /goal don’t stop until digest is
  posted and STATE.md is updated.

▲ Claude
  Creating routine: nightly-eval-compounding
  - model: claude-fable-5
  - harness: claude managed agent (sandbox)
  - trigger: schedule (0 7 * * *)
  - grader: independent Haiku sub-agent (Outcomes)
✓ Active. First run tomorrow 07:00 local. Skill set will compound.

PART 3 · The Self-Improvement Layer 第 3 部分 · 自我改进层

10. The 5-stage memory progression. 10. 五阶段记忆演进

The single most useful framing for what “agent memory” means in practice comes from the Anthropic team’s Continual Learning Bench 1.0 experiment. Effective use of memory requires a progression of five stages. Each stage is a structural move; each model exits the progression at a different point. “智能体记忆”在实际中最有用的框架，源自Anthropic团队的持续学习基准1.0实验。有效利用记忆需经历五个阶段，每个阶段都是一次结构性推进，不同模型会在不同节点退出该进程。

• 1. Fail - the agent gets something wrong and documents the failure with enough detail to be useful later.1. 失败——智能体出现错误，并记录失败细节以备后续参考。
• 2. Investigate — before moving on, the agent figures out why the failure happened.2. 调查——在继续推进前，智能体需查明失败原因。
• 3. Verify - the agent turns the diagnosis into a checked fact, not a guess.3. 验证——智能体将诊断结果转化为经核验的事实，而非主观猜测。
• 4. Distill - the agent turns the verification into a general rule that applies beyond the specific case.4. 提炼——智能体将验证结果转化为通用规则，使其能适用于具体案例之外的情况。
• 5. Consult - on the next task, the agent reads the rule instead of re-deriving the fact from scratch.5. 援引——在后续任务中，智能体直接读取该规则，无需从头推导这一事实。

The measured difference between models on a SQL exploration task from the Continual Learning Bench, each model with memory provided: 在持续学习基准测试的SQL探索任务中，各配备记忆模块的模型之间测得的差异如下：

• Sonnet 4.6 exits at step 1. Its memory store is a list of failure notes and open guesses (“maybe prc instead of prc_usd?”). It rarely consults prior notes. Memory exists but doesn’t compound.Sonnet 4.6在第一步就终止了。其记忆存储仅是一份错误记录和未决猜想清单（例如”或许是prc而非prc_usd？”）。它很少查阅之前的记录。记忆虽然存在，却未能形成累积效应。
• Opus 4.7 exits at step 3. It creates a schema reference with uncertainty flagged (“possibly prc in cents? Verify.”). Verification coverage runs 7–33% (median ~17%) of questions.Opus 4.7 在第三步退出。它创建了一个带有不确定性标记的模式引用（“可能是以美分计价的定价？需核实。”）。验证覆盖率覆盖了 7%–33%（中位数约 17%）的问题。
• Fable 5 tends to complete the progression. In its strongest runs, verification coverage reaches 73% (22 of 30), and it distills learnings into general rules that help with future tasks.Fable 5 倾向于完成整个流程。在其最强运行中，验证覆盖率达到 73%（30 题中答对 22 题），并将经验提炼为通用规则，帮助处理未来任务。

11. The state file. Where memory actually lives.11. 状态文件——记忆真正存在的地方。

The 5-stage progression is the mental model. The state file is where the model writes each stage’s output. For Fable 5 running in Claude Managed Agents, memory is a mounted filesystem that survives between sessions; in Claude Code locally, a markdown file or a Linear board does the same job. 五阶段流程是思维模型。状态文件则是模型记录每个阶段输出的地方。对于在 Claude 托管代理中运行的 Fable 5，记忆是挂载在会期间的持久化文件系统；而在本地 Claude Code 中，Markdown 文件或 Linear 看板也能起到相同作用。

The structure of a state file that actually supports the 5-stage progression: 支持五阶段进展的状态文件结构如下：

# Project memory · trading-platform

## Verified facts # stage 3 — stop guessing about these
- prc is in dollars, not cents. Verified via SELECT MIN(prc), MAX(prc) FROM trades.
- user_id matches auth_users.uid via JOIN, not auth_users.id. Confirmed 2026-06-09.
- Test database uses Stripe sandbox keys; production uses real keys via env.

## General rules # stage 4 — consult before re-deriving
- When querying time-bucketed metrics, always include timezone (default UTC mismatches).
- Auth middleware order matters: rate_limit -> jwt -> rbac. Reversing causes 401s.
- For migrations, never use ALTER on tables >1M rows without batching.

## Open failures (investigate next session) # stage 1 → 2
- 2026-06-09: tests/e2e/checkout flakes ~1 in 50 runs. Hypothesis: webhook race.
  Reproduction steps in debug/checkout-flake.md.

## Lessons learned # stage 4 distillations
- PowerShell hits TLS 1.2 issue on Windows CI runners. Always shell out to bash.
- Stripe webhook tests require STRIPE_WEBHOOK_SECRET. Skip with clear message if missing.

## Last session # stage 5 — resume, don’t restart
2026-06-10 03:30 UTC · 7 failures classified, 3 fixes drafted (claude/fix-*), 4 escalated.
Next: verify the auth middleware fix in claude/fix-rate-limit-order against production load.

The file has five sections matching the five stages. Verified facts is stage 3 output - things the agent stopped guessing about. General rules is stage 4 - distilled rules that apply beyond the specific case. Open failures is stages 1–2 work in progress. Lessons learned is more stage 4 output. 该文件包含与五个阶段相对应的五个部分。已验证事实是第三阶段的输出——智能体不再猜测的内容。通用规则是第四阶段的输出——适用于超越特定案例的提炼规则。未解决失败是第一、二阶段的进行中工作。经验教训是更多第四阶段的输出。

Last session is the resume pointer for stage 5.最后会话是第五阶段的恢复指针。

Two operational rules that decide whether this file actually compounds or just grows: 两条操作规则决定了这个文件是真正实现复利增长，还是仅仅线性累积：

• Write before walking away. Every Fable 5 session ends by updating STATE.md - what was tried, what passed, what failed, what new rules survived. If the session doesn’t finish with a write, the next one restarts from zero.离开前必须记录。每个 Fable 5 会话结束时必须更新 STATE.md——记录尝试过什么、哪些成功、哪些失败、哪些新规则得以保留。如果会话结束时没有完成记录，下一次会话将从零开始。
• Read at session start. Every new session begins by reading STATE.md and the most relevant Skills. The Continual Learning Bench data shows that without this, Sonnet-class memory behavior shows up even in Fable 5.会话开始时必须阅读。每个新会话开始时都要阅读 STATE.md 和最相关的技能资料。《持续学习基准》数据显示，如果不这样做，即使使用 Fable 5，也会出现类似 Sonnet 的记忆行为问题。

12. Skills that compound. Write the lesson into the Skill, not just the chat.12. 可复利积累的技能。将经验教训写进技能模块，而不只是留在聊天记录中。

STATE.md is for project memory. Skills are for procedural memory - the “how to do this kind of thing” that should apply across projects. STATE.md 用于项目记忆。技能则属于程序性记忆——即“如何完成这类任务”的知识，应该适用于跨项目场景。

The compounding pattern: after any non-trivial failure, write the lesson into the Skill itself. The Skill gets sharper every time the system runs.复合积累模式：在每次非琐碎失败后，将教训直接写入技能本身。每当系统运行一次，该技能就会变得更精准。

A Skill that’s been compounding for two weeks looks different from a fresh one. New sections appear: known failure modes, rules that came out of post-mortems, anti-patterns observed in production. 经过两周复合积累的技能与新技能截然不同。新增的章节包括：已知的失败模式、事后复盘总结出的规则、以及在生产环境中观察到的反模式。

The Skill is no longer a static set of instructions; it’s an accumulating record of what the team has actually learned. 技能不再是一套静态的指令集，而是团队实际经验教训的累积记录。

---
name: ci-triage
description: Classify CI failures, draft fixes for easy ones, escalate the rest.
  Trigger on workflow_run.failure or on the morning triage routine.
---

# CI triage skill

## Classification rules
- env: missing secret, wrong env var. # escalate to human, never auto-fix
- flake: passes on retry without code change. # retry once, then file
- bug: deterministic failure tied to recent commit. # draft fix
- dependency: tied to version bump. # draft rollback
- infra: timeout, OOM, runner issue. # escalate

## Known failure modes # added by the loop over 14 days
- webhook-race: e2e checkout flakes when Stripe webhook arrives mid-test.
  Fix: add 2s settle delay in tests/utils/webhook.ts.
- tls-handshake: Windows runners fail TLS 1.2 in PowerShell. Use bash.
- db-migration: ALTER on trades table >1M rows times out at 30s. Batch in 10k chunks.

## Anti-patterns (do NOT do) # added after real incidents
- Never disable a failing test to make CI green. File it instead.
- Never modify .github/workflows/ without human approval.
- Never touch src/payments/ or src/billing/ without security review.

## State
Update STATE.md after each run with classifications, fixes drafted, escalations.

## Eval suite # step 13 — the loop verifies the skill
Run against eval/ci-triage-cases.jsonl weekly. Any newly-failing case →
add to known failure modes after Outcomes verifier confirms.

The compounding contract: every confirmed lesson goes into a Skill, not just STATE.md. STATE.md is project-scoped and dies with the project. Skills live in ~/.claude/skills/ and travel with you. 复合契约：每一个被验证的经验教训都会沉淀为一项技能，而非仅记录在 STATE.md 中。STATE.md 的作用范围仅限于当前项目，并随项目终结而失效。技能则存储于 ~/.claude/skills/ 目录下，可随身携带。

Two weeks of disciplined writing produces a Skill that materially outperforms whatever Fable 5 would derive from scratch on a fresh project. 经过两周的持续记录与沉淀所形成的技能，在性能上实质性地优于 Fable 5 在全新项目中从零推导出的结果。

13. Self-verification via vision. Fable 5 checks its own UI against the goal.13. 基于视觉的自我验证。Fable 5 通过比对自身用户界面与目标状态进行校验。 ```

One of the headline capabilities Anthropic ships with Fable 5 is “uses vision to check outputs against goals.” This sounds abstract until you see what it actually replaces: the human eyeballing a screenshot to confirm the UI looks right. Anthropic 在 Fable 5 中推出的重要能力之一是“通过视觉检查输出是否达成目标”。这种表述听起来可能有些抽象，直到你看到它实际替代了什么：人工肉眼检查截图以确认界面效果无误。

Fable 5 does that step itself, in the loop, before declaring done. Fable 5 在循环中自主完成这一步，然后才宣告任务完成。

The pattern in production: 在生产环境中的模式如下：

• Maker sub-agent writes the UI code. Renders the result to a screenshot.制作者子代理编写界面代码，并将结果渲染为截图。
• Verifier sub-agent reads the screenshot with vision, compares it against the goal description, against design tokens in the project Skill, and against the previous screenshot from STATE.md.验证子代理通过视觉读取截图，将其与目标描述、项目技能中的设计令牌以及STATE.md中的前一次截图进行比对。
• Verdict goes back to the loop. Match → mark task complete. Mismatch → describe the gap, hand back to maker with a structured diff.评估结果返回循环。匹配→标记任务完成。不匹配→描述差异，将带有结构化差异信息的反馈返回给制造者。

This pattern is what Anthropic measured in the Parameter Golf experiment under the same harness: Fable 5 looked at training charts (visual artifact) and decided whether the curve matched the criterion. 这种模式正是Anthropic在参数高尔夫实验中通过相同测试框架所衡量的：Fable 5查看了训练图表（视觉工件），并判断曲线是否符合标准。

No human in the loop reading the chart. The verifier read the chart. 无需人类在循环中查看图表——验证器代劳了。

14. The Mythos safety boundary. What Fable 5 won’t do, and how to design around it.14. 神话安全边界。Fable 5 不会做什么，以及如何围绕它进行设计。

he last step is the one most easily skipped on day one and most expensive to learn the hard way. 最后一步最容易在第一天被忽略，但通过惨痛教训来学习的代价也最高。

Fable 5 ships with built-in safety classifiers that decline to respond in specific high-risk domains - cybersecurity vulnerability research, biology, chemistry, and model distillation. In those domains, Anthropic falls Fable 5 back to Claude Opus 4.8 automatically. This is documented; it’s not a bug.Fable 5 内置了安全分类器，在某些高风险领域（如网络安全漏洞研究、生物学、化学和模型蒸馏）会拒绝响应。在这些领域，Anthropic 会自动将 Fable 5 回退到 Claude Opus 4.8。这一机制有文档记录，并非系统缺陷。

What this means for a self-improving system that runs autonomously: 这对自主运行的自改进系统意味着什么：

• If your system touches security tooling (SAST scans, exploit research, penetration testing logic, even some classes of code review), expect classifier blocks. Architect for the fallback: route those tasks to Opus 4.8 explicitly, or surface the block to a human reviewer.如果你的系统涉及安全工具（如SAST扫描、漏洞研究、渗透测试逻辑，甚至某些代码审查类别），预计会受到分类器拦截。为应对这种降级情况设计架构：明确将这些任务路由至Opus 4.8，或将其提交给人工审核人员。
• Same for biology, chemistry, and distillation domains. The classifier is broad. A scientific computing workflow might trigger it; a code review of crypto primitives might trigger it.生物学、化学和蒸馏领域也是如此。该分类器的覆盖范围很广。科学计算工作流程可能触发它，加密原语的代码审查也可能触发它。
• Design your Skills to surface the fallback gracefully. A Skill should know which kinds of tasks it produces that may hit the classifier and document the expected behavior. A loop that silently fails on a classifier block looks identical to a loop that fails on a real error — until you debug it.设计你的技能模块以优雅地呈现降级处理方案。每个技能模块应了解其生成的哪些任务类型可能触碰分类器，并记录预期行为。一个在分类器拦截时静默失败的循环，与其在真实错误时失败的循环看起来完全一样——直到你进行调试。
• Audit the system card. Fable 5’s 319-page system card documents the classifier’s scope. The launch generated controversy in mid-June 2026 because some downgrade behaviors were discovered buried in the document. Read it before deploying to production.审查系统说明书。Fable 5的319页系统说明书详细记录了分类器的覆盖范围。2026年6月中旬的发布曾引发争议，因为文档深处隐藏的某些降级行为被曝光。在生产环境部署前，请务必仔细阅读。

The general design principle: treat the safety boundary as a known fallback, not as a failure mode. A self-improving system that ships with explicit handling of the boundary stays robust as the classifier evolves. A system that ignores it produces silent regressions when Anthropic updates the policy.总体设计原则：将安全边界视为已知的兜底机制，而非故障模式。一个自带边界显式处理机制的自进化系统，能在分类器演进过程中保持鲁棒性。而忽略该边界的系统，当Anthropic更新策略时会产生静默退化。

§ The mistakes that keep Fable 5 at 10% of its potential § 让Fable 5仅发挥10%潜能的常见误区

• Using Fable 5 like Sonnet 4.6 with more context. A 5-minute prompt-and-close session burns Mythos-tier pricing for no compound effect.把Fable 5当作拥有更大上下文的Sonnet 4.6来使用。五分钟的提示词输入与关闭操作耗费着Mythos级别的费用，却无法产生复合效应。
• Self-critique instead of an independent verifier. The maker grades its own homework. Anthropic measured the difference; the team explicitly documents the verifier sub-agent pattern.用自我批评取代独立验证器。创作者为自己的作业打分。Anthropic已衡量过两者的差异；团队明确文档化了验证器子代理模式。
• No STATE.md. Every session restarts from zero. The Continual Learning Bench data shows this is where 70%+ of Fable 5’s memory advantage disappears.无 STATE.md 文件。每次会话都从零开始。持续学习基准数据显示，这正是 Fable 5 超过 70% 记忆优势消失的原因。
• Skills that never get written to. A static Skill is fine; a Skill that doesn’t accumulate lessons after real failures is wasted scaffolding.从未被写入的技能。静态技能本身没问题；但一个在经历真实失败后未能积累经验教训的技能，不过是浪费资源的空架子。
• Fable 5 on tasks Sonnet 4.6 would handle. Doc updates, simple refactors, lint fixes. Route by complexity; reserve Fable 5 for the orchestrator role.Fable 5 处理原本由 Sonnet 4.6 承担的任务：文档更新、简单重构、代码规范修复。按复杂度分流，将 Fable 5 保留给编排者角色。
• Running long sessions on a laptop. Days-long capability requires cloud infrastructure (CMA or Routines). A closed laptop kills the session.在笔记本电脑上运行长会话。需要云端基础设施（CMA 或例程）才能实现持续数天的能力。合上笔记本就会终止会话。
• Ignoring the Mythos safety boundary. Classifier blocks on cyber/bio/chem produce silent regressions. Architect for the fallback explicitly.忽略 Mythos 安全边界。针对网络/生物/化学的分类器模块产生静默退化。明确为备用方案设计架构。
• No vision-verify on visual tasks. UI, dashboards, design fidelity — checking these with text-only verifiers misses the failure mode that matters.无视觉验证的视觉任务检查。UI、仪表盘、设计保真度——仅用纯文本验证器检查会遗漏关键失效模式。
• Skipping /goal or Outcomes. Without an objective stop condition checked by an independent grader, loops stop at “handled enough” instead of done.跳过 /goal 或 Outcomes。若缺乏由独立评分器检查的目标终止条件，循环会在”处理得差不多”而非”完成”时停止。
• No retention policy review. Sensitive data through a Fable 5 routine without checking the 30-day / 2-year terms creates compliance issues silently.无保留策略审查。敏感数据通过 Fable 5 例程流转却未检查 30 天/2 年期限，会静默产生合规问题。

Conclusion: 结论：

Fable 5 isn’t a faster chat tool. It’s the substrate for a system that compounds.Fable 5 并非一个更快的聊天工具。它是一个具备自我进化能力的系统的基础架构。

The first publicly available Mythos-class model didn’t ship to be prompted faster. It shipped to be the orchestrator of a self-improving system you build around it. 首个面向公众的 Mythos 级模型推出，并非为了提供更快捷的提示响应，而是为了成为您围绕其构建的自我进化系统的中枢调度器。

The capability headlines - days-long sessions, sub-agent delegation, vision self-check, accumulated memory - only earn their pricing if the system around the model is doing its job. 其功能亮点——持续数日的会话、子代理委派、视觉自检机制、累积记忆——只有在围绕该模型构建的整个系统有效运作时，才能体现其价值。

The Anthropic team’s own experiments make the gap visible. Parameter Golf: Fable 5 with an independent verifier explored larger architectural changes and pushed through negative intermediate results to land ~6× more improvement than Opus 4.7. Anthropic团队自己的实验让差距清晰可见。参数高尔夫：Fable 5通过独立验证器探索了更大范围的架构变更，并推动负面中间结果，最终比Opus 4.7实现了约6倍的改进。

Continual Learning Bench: Fable 5 with memory completed the full 5-stage progression with 73% verification coverage, against Opus 4.7’s 17%. The model is the same in both halves of every comparison. The system around it is what changed.持续学习基准：Fable 5通过记忆功能完成了完整的5阶段进阶，验证覆盖率达73%，而Opus 4.7仅为17%。模型在同一对比的两端完全一致，改变的是其周围的系统。

Pick one layer of the compound stack you weren’t doing - probably the verifier sub-agent (step 6), the state file (step 11), or vision-verify (step 13) - and add it tomorrow. Then the next. 从复合栈中选取一个你还未实践的层级——可能是验证子智能体（步骤6）、状态文件（步骤11）或视觉验证（步骤13）——明天就将其加入。接着再添加下一个。

Self-improvement is a property of the system, not the model. Build the system.自我提升是系统的属性，而非模型的属性。构建系统。