# Designing loops with Fable 5
**作者**: Lance Martin
**日期**: 2026-05-04T17:59:12.000Z
**来源**: [https://x.com/RLanceMartin/status/2064397389189071163](https://x.com/RLanceMartin/status/2064397389189071163)
---
Mythos-class models like Claude Fable 5 have changed the way many of us work at Anthropic. I want to share two tips for getting the most out of this class of models.
Self-correction loops
There’s been a lot of interest in loops recently. @bcherny has mentioned that “(his) job is to write loops.” Letting models hillclimb on an evaluation is a common recipe for improving task performance: /goal in Claude Code and Outcomes in Claude Managed Agent are primitives that let you apply this general recipe for your specific task.
As mentioned in our prompting guide, Fable 5 is good at self-correcting in a loop. A well designed goal or rubric adds feedback to the environment that Claude is running in. This let’s Claude run, collect feedback via the goal or rubric, self-correct, and proceed until the goal or rubric is satisfied.
I’ll share one toy example that I used to test Fable: Parameter Golf is an open source ML engineering challenge to train the best model that fits in a 16MB artifact in < 10 minutes on 8xH100s.
It’s a bit like @karpathy's autoresearch project: it tests the ability of an agent to edit basic training code (a single train_gpt.py file), launch training, poll the log, read the score, and decide what experiment to run next.
I compared Fable 5 to Opus 4.7 on this challenge using Claude Managed Agents (CMA). CMA provides the agent harness as well as a hosted sandbox, so it’s well-suited for long-running tasks with Fable 5. For Parameter Golf, I gave CMA access to 8xH100 GPUs as a self-hosted sandbox.
One subtle point: what does the judging is important. We’ve seen that models have problems with self-critique on their own outputs. Prithvi Rajasekaran wrote about this in our engineering blog here.
We’ve found that a verifier sub-agent tends to outperform self-critique with Fable 5, because grading is done in an independent context window. Outcomes in CMA handles this by spawning a grader sub-agent for you.
For each test, I supplied a rubric (a file) with the nine checkable criteria (e.g., run a baseline, run 20 experiments, etc). Then, I ran Parameter Golf for up to 8 hours. The Outcomes grader confirmed that all experimental criteria were met before allowing Claude to stop the work.
Fable 5 improved the training pipeline ~6x more than Opus 4.7. If we consider experiments as structural (e.g., architecture changes) or scalar (e.g., adjusts a constant), Fable 5 bet on larger structural changes and showed resilience (e.g., pushing through a quantization regression to its biggest win).
Opus 4.7's first experiment produced a small win and nearly everything after followed the same template: adjust a scalar, measure, keep if positive.
Memory
Memory is another area where Fable excels. We can think about this as a outer loop that spans across sessions: Claude writes to memory during a session and those memories can be retrieved in future sessions.
@pgasawa and team recently published Continual Learning Bench 1.0, so I wanted to test this on Fable 5 vs earlier models.
> **Parth Asawa@pgasawa**: [原文链接](https://x.com/pgasawa/status/2051361012838957144)
>
> Today, we’re releasing Continual Learning Bench 1.0: the first, realistic benchmark for measuring how AI systems can improve in online settings.
> Benchmarks today assume models are stateless. Each example is independent, and once a system finishes a task, it moves on as if
>
> 
I compared Fable 5, Opus 4.7, and Sonnet 4.6 on one of the tasks from the benchmark: the task asks an agent to answer sequential questions given access to a SQL database. Each question is a separate agent session and memory is provided.
For this, I used CMA with memory, which gives each agent access to a mounted filesystem that can be shared across sessions.
For this task, effective use of memory benefits from a progression: fail (get something wrong and document), investigate (before moving on, figure out why), verify (turn the diagnosis into a checked fact), distill (turn verification into a general rule), and consult (read the rule, instead of re-deriving it).
Sonnet 4.6 exits around step 1: its store is a list of failure notes and open guesses (e.g., "maybe prc instead of prc_usd?"). It rarely consults prior notes. To improve performance, task-specific memory instructions are needed.
Opus 4.7 exits around step 3: it creates a schema reference with uncertainty flagged (e.g., "possibly prc in cents? Verify."), but verification coverage is low: at 7-33% of questions (median run ~17%).
Fable 5 tends to complete the progression: in its strongest runs, verification coverage is up to 73% (22 of 30) and it distills learnings into general rules that help with future tasks.
Rather than directly prompting and steering Fable 5, it's often better to design loops that let the model to self-correct in response to environment feedback (e.g., /goal or Outcomes) and manage its own context (e.g., via memory).
I've shared just a few small scale experiments that I've run, but its worth testing Fable 5 for yourself on challenging tasks and using loops for self-correction or memory.
To get started, see our docs or ask the latest version of Claude Code, which can use our built-in /claude-api skill to tell you about Fable 5 (e.g., prompting best practices), /goal, Claude Managed Agents, or other API features.
## 相关链接
- [Lance Martin](https://x.com/RLanceMartin)
- [@RLanceMartin](https://x.com/RLanceMartin)
- [589K](https://x.com/RLanceMartin/status/2064397389189071163/analytics)
- [@bcherny](https://x.com/@bcherny)
- [has mentioned](https://x.com/sairahul1/status/2064279904989147577?s=20)
- [/goal](https://code.claude.com/docs/en/goal)
- [Outcomes](https://platform.claude.com/docs/en/managed-agents/define-outcomes)
- [prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5)
- [Parameter Golf](https://github.com/openai/parameter-golf)
- [@karpathy](https://x.com/@karpathy)
- [autoresearch](https://github.com/karpathy/autoresearch)
- [Claude Managed Agents](https://platform.claude.com/docs/en/managed-agents/overview)
- [the agent harness as well as a hosted sandbox](https://www.anthropic.com/engineering/managed-agents)
- [self-hosted sandbox](https://platform.claude.com/docs/en/managed-agents/self-hosted-sandboxes)
- [here](https://www.anthropic.com/engineering/harness-design-long-running-apps)
- [tends to outperform](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5)
- [Outcomes](https://platform.claude.com/docs/en/managed-agents/define-outcomes)
- [another area where Fable excels](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices)
- [@pgasawa](https://x.com/@pgasawa)
- [May 5](https://x.com/pgasawa/status/2051361012838957144)
- [831K](https://x.com/pgasawa/status/2051361012838957144/analytics)
- [memory](https://platform.claude.com/docs/en/managed-agents/memory)
- [docs](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5)
- [/claude-api](https://github.com/anthropics/skills/tree/main/skills/claude-api)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [1:21 AM · Jun 10, 2026](https://x.com/RLanceMartin/status/2064397389189071163)
- [589.8K Views](https://x.com/RLanceMartin/status/2064397389189071163/analytics)
- [View quotes](https://x.com/RLanceMartin/status/2064397389189071163/quotes)
---
*导出时间: 2026/6/10 12:46:23*
---
## 中文翻译
# 使用 Fable 5 设计循环
**作者**: Lance Martin
**日期**: 2026-05-04T17:59:12.000Z
**来源**: [https://x.com/RLanceMartin/status/2064397389189071163](https://x.com/RLanceMartin/status/2064397389189071163)
---
像 Claude Fable 5 这样的 Mythos 级模型改变了许多人在 Anthropic 的工作方式。我想分享两个技巧，帮助你充分发挥这类模型的作用。
自纠正循环
最近大家对“循环”的兴趣很浓。@bcherny 曾提到“（他的）工作就是写循环”。让模型在评估中进行“爬山”是提高任务性能的常见配方：Claude Code 中的 /goal 和 Claude 托管代理中的 Outcomes 是原语，让你能将这一通用配方应用到具体任务中。
正如我们的提示词指南所提到的，Fable 5 擅长在循环中进行自纠正。一个设计良好的目标或评价标准能为 Claude 所处的环境提供反馈。这让 Claude 得以运行，通过目标或评价标准收集反馈，进行自纠正，并持续执行直到目标或评价标准得到满足。
我想分享一个用来测试 Fable 的简单示例：Parameter Golf 是一个开源的机器学习工程挑战，要求在 8xH100 上于 10 分钟内训练出最佳模型，且模型大小需适配在 16MB 的产物中。
这有点像 @karpathy 的 autoresearch 项目：它测试一个代理编辑基础训练代码（一个单独的 train_gpt.py 文件）、启动训练、轮询日志、读取分数并决定下一个实验是什么的能力。
我在这个挑战上使用 Claude 托管代理（CMA）对比了 Fable 5 和 Opus 4.7。CMA 提供了代理的框架（harness）以及一个托管的沙箱，因此非常适合让 Fable 5 运行长时间任务。对于 Parameter Golf，我让 CMA 接入了 8xH100 GPU 作为自托管沙箱。
一个微妙之处：由什么来进行评判很重要。我们发现模型在对自己产生的输出进行自我评判时会遇到问题。Prithvi Rajasekaran 在我们的工程博客中写过这一点。
我们发现，对于 Fable 5，使用一个验证者子代理往往胜过自评自纠，因为评分是在独立的上下文窗口中完成的。CMA 中的 Outcomes 功能通过为你生成一个评分子代理来处理这一点。
对于每次测试，我提供了一份评价标准（一个文件），其中包含九项可核查的标准（例如：运行基线、运行 20 个实验等）。然后，我让 Parameter Golf 运行长达 8 小时。Outcomes 评分者会确认所有实验标准均已满足，然后才允许 Claude 停止工作。
Fable 5 对训练流程的改进幅度比 Opus 4.7 高出约 6 倍。如果我们将实验分为结构性（例如架构变更）或标量性（例如调整某个常数），Fable 5 更倾向于进行较大的结构性变更，并展现出韧性（例如，在经历量化导致的性能回退后仍坚持，最终取得最大胜利）。
Opus 4.7 的第一个实验带来了一个小幅提升，随后几乎所有的实验都遵循同一模板：调整一个标量、测量、如果效果正向则保留。
记忆
记忆是 Fable 擅长的另一个领域。我们可以将其视为一个跨会话的外层循环：Claude 在会话期间写入记忆，而这些记忆可以在未来的会话中被检索。
@pgasawa 及其团队最近发布了 Continual Learning Bench 1.0，所以我想在 Fable 5 和早期模型上测试这一点。
> **Parth Asawa@pgasawa**: [原文链接](https://x.com/pgasawa/status/2051361012838957144)
>
> 今天，我们发布了 Continual Learning Bench 1.0：首个用于衡量 AI 系统在在线设置中如何改进的现实基准。
> 今天的基准假设模型是无状态的。每个示例都是独立的，一旦系统完成一个任务，它就会继续前进，仿佛……
>
> 
我在基准测试的一项任务上对比了 Fable 5、Opus 4.7 和 Sonnet 4.6：该任务要求一个代理在可以访问 SQL 数据库的情况下回答一系列问题。每个问题都是一个独立的代理会话，并提供了记忆功能。
为此，我使用了带记忆功能的 CMA，它为每个代理提供了一个可跨会话共享的挂载文件系统。
对于这项任务，有效利用记忆得益于一个循序渐进的过程：失败（把事情做错并记录下来）、调查（在继续之前弄清原因）、验证（将诊断转化为已核查的事实）、提炼（将验证转化为一般规则）、以及查阅（阅读规则，而不是重新推导）。
Sonnet 4.6 通常止步于第 1 步：它的存储只是一列失败笔记和未解的猜测（例如，“也许是用 prc 而不是 prc_usd？”）。它很少查阅之前的笔记。要提升性能，需要针对特定任务的记忆指令。
Opus 4.7 通常止步于第 3 步：它会创建一个标注了不确定性的模式参考（例如，“可能 prc 是以分为单位？验证。”），但验证覆盖率很低：仅覆盖 7-33% 的问题（中位数运行约为 17%）。
Fable 5 倾向于完成整个过程：在其表现最好的运行中，验证覆盖率高达 73%（30 个问题中有 22 个），并且它能将所学提炼为有助于未来任务的一般规则。
与其直接提示和引导 Fable 5，通常不如设计一些循环，让模型能够根据环境反馈（例如 /goal 或 Outcomes）进行自纠正，并管理其自身的上下文（例如通过记忆）。
我只分享了一些我做过的小规模实验，但值得自己在具有挑战性的任务上测试 Fable 5，并利用循环进行自纠正或记忆管理。
要开始使用，请查看我们的文档或询问最新版本的 Claude Code，它可以使用我们内置的 /claude-api 技能来告诉你有关 Fable 5 的信息（例如提示词最佳实践）、/goal、Claude 托管代理或其他 API 功能。
## 相关链接
- [Lance Martin](https://x.com/RLanceMartin)
- [@RLanceMartin](https://x.com/RLanceMartin)
- [589K](https://x.com/RLanceMartin/status/2064397389189071163/analytics)
- [@bcherny](https://x.com/@bcherny)
- [提到过](https://x.com/sairahul1/status/2064279904989147577?s=20)
- [/goal](https://code.claude.com/docs/en/goal)
- [Outcomes](https://platform.claude.com/docs/en/managed-agents/define-outcomes)
- [提示词指南](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5)
- [Parameter Golf](https://github.com/openai/parameter-golf)
- [@karpathy](https://x.com/@karpathy)
- [autoresearch](https://github.com/karpathy/autoresearch)
- [Claude 托管代理](https://platform.claude.com/docs/en/managed-agents/overview)
- [代理框架以及托管沙箱](https://www.anthropic.com/engineering/managed-agents)
- [自托管沙箱](https://platform.claude.com/docs/en/managed-agents/self-hosted-sandboxes)
- [这里](https://www.anthropic.com/engineering/harness-design-long-running-apps)
- [往往胜过](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5)
- [Outcomes](https://platform.claude.com/docs/en/managed-agents/define-outcomes)
- [Fable 擅长的另一个领域](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices)
- [@pgasawa](https://x.com/@pgasawa)
- [5月5日](https://x.com/pgasawa/status/2051361012838957144)
- [831K](https://x.com/pgasawa/status/2051361012838957144/analytics)
- [记忆](https://platform.claude.com/docs/en/managed-agents/memory)
- [文档](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5)
- [/claude-api](https://github.com/anthropics/skills/tree/main/skills/claude-api)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [1:21 AM · Jun 10, 2026](https://x.com/RLanceMartin/status/2064397389189071163)
- [589.8K 次观看](https://x.com/RLanceMartin/status/2064397389189071163/analytics)
- [查看引用](https://x.com/RLanceMartin/status/2064397389189071163/quotes)
---
*导出时间: 2026/6/10 12:46:23*