32 分钟演讲 · 15 个核心判断

当 AI Agent 学会做梦

Lamis Mukta · Anthropic · AI Native DevCon June 2026
YouTube · 完整笔记 →

🧠 Lamis Mukta
🏢 Anthropic Applied AI
🎓 Oxford 数学硕士
💤 Dreaming 架构师
🔥 最炸裂的判断
1
犀利
"Intelligence alone is not going to compound. Without memory, task 50 is no better than task 1."
Intelligence alone doesn't compound
这不是能力问题,是架构问题。模型变聪明了,但如果没有 Context Engineering,你的 Agent 永远不会从经验中成长。
2
犀利
"CLAUDE.md was unreasonably effective — a markdown file that just gives the agent a couple of instructions was so good at steering it."
一个 Markdown 文件的效果好到不合理。这是 Anthropic 整条记忆进化路线的起点——简单的东西先行,复杂的方案后补。Do the simple thing that works.
3
犀利
"Agents are actually just very good at using Bash and Grep. Just let them search over the file system."
Anthropic 当前最佳实践:别设计专门的记忆 API,把记忆放进 Markdown 文件,让 Agent 用标准工具自己搜。文件系统就是最好的记忆系统。
4
犀利
"All of this stuff has only happened in this past year. This is very much an open area of research."
从 CLAUDE.md 到 Dreaming,整条进化路线只花了一年。领域远未成熟,最佳实践还在剧烈演化中。
🏗️ 架构洞察
5
架构
"Skills: it's as if I had a bookshelf — I can scan the titles and pick one off the shelf when I need to."
How agent memory evolved
渐进式披露(Progressive Disclosure)的精髓:Agent 只看书脊标题,不把每本书背下来。Skill 顶部的 front matter 就是书脊——几行字就能决定要不要展开。
6
对比

带内记忆 In-band

  • 会话中读写
  • 资源与任务共享
  • 单 session 视野
  • 下次立即生效

Dreaming Out-of-band

  • 会话间异步运行
  • 独立 token 预算
  • 跨所有 Agent 视野
  • 下一轮批处理后生效
两者不是替代关系,而是双轮驱动。带内负责即时反馈,Dreaming 负责全局模式识别。
7
工程
"When an agent wants to write, it takes a hash before and after. If they don't match, the write is rejected."
本质就是乐观锁。Agent 更新记忆时先拍快照,写完再校验。hash 不匹配就重试。多 Agent 并发写记忆的工程解法。
8
架构
"Dreaming does not change model weights — the system improves even if the model stays the same."
How dreaming works
Dreaming 是外部记忆层的整理流程,不是微调或 RLHF。不需要 GPU,不需要训练基础设施,随时可以回滚。
🛡️ 生产实践
9
工程
"You wouldn't want one agent to just decide it should update the organization-wide context. That should be read-only."
Built for multi-agent systems
记忆需要权限分层:组织级只读、项目级受限写入、Agent 个人级完全读写。一个 Agent 的错误判断不该扩散到整个组织。
10
工程
"We're not just looking at the back-and-forth — we're also really scrutinizing tool calls and all the metadata."
Dreaming 分析 transcript 时不只看对话内容——工具调用、技能使用、错误日志等元数据同等重要。工具配置出错只有通过元数据才能发现。
11
实用
"You have the ability to steer how dreaming agents go about the problem — tell them what's important and what's not."
Dreaming 不是黑盒。你可以定向引导:告诉整理 Agent 你的场景里哪类模式值得关注、哪类可以忽略。
📊 数据 & 案例
12
数据
"Harvey's task completion rates rose roughly 6x in internal testing once Dreaming was turned on."
法律 AI 公司 Harvey 的案例:Agent 之前在 session 间反复忘记文件格式坑和工具变通方法,同样的任务反复失败。Dreaming 持久化了这些经验,打破了失败循环。Anthropic 内部也看到 +10 百分点任务成功率、+8.4% docx 质量提升。
13
成本
"It sounds expensive — why chuck extra resources at this? But agents one-shot things more effectively, so costs go down."
Dreaming 看似额外开销,实际上是投资:Agent 变聪明后,执行任务的 token 消耗和重试次数减少,总成本反而下降。
💬 全场最佳
14
犀利
"At which point are we reinventing databases from first principles again?"
全场最佳提问。版本控制、并发锁、权限系统——这些不就是数据库几十年前解决的问题吗?Lamis 坦承"确实在回归这些实践",但强调路径是:先让 Agent 自由探索,观察什么有效,再把验证过的模式硬编码回 harness。
15
实用
"This is not coding specific — I use memory for presentations, writing style, slide preferences. It develops over time."
Keep thinking, keep learning, keep dreaming
记忆系统的价值远不止代码。Lamis 自己就用它来记住写作风格、演示偏好。任何需要"下次做得更好"的场景都适用。

Context Engineering 是乘法效应,
文件系统是最好的记忆系统,
带内记忆有天花板,
Dreaming 在睡眠中让 Agent 持续进化。

Keep thinking, keep learning, keep dreaming.