Manus

Manus：AI agent 公司，context engineering 五维框架和 prefix cache 优化实践的主要贡献者

实

ENTITY · MANUS · GENERAL AUTONOMOUS AGENT · 5-D CONTEXT ENGINEERING · PRODUCTION EXPERIENCE

Manus

AI agent company — a production-validated context engineering framework evolved through “Stochastic Graduate Descent” iteration

Manus went through four architectural rebuilds and distilled a five-dimensional context engineering framework. Core economic insight: agent loops run prefill/decode at roughly 100:1, making prefix cache hit rate the primary production cost lever (Claude Sonnet: $0.30/MTok hit vs.$ 3.00/MTok miss — a 10x cost gap).

5-D Context Engineering Framework

OffloadingFile system as unbounded external memory, zero token cost — move content “out of” context

ReductionOnly recoverable compression (keep path/URL pointers) — prevent permanent information loss

RetrievalFile search tools as a structured retrieval layer — load on demand rather than all at once

IsolationSub-agents run in separate context windows — isolation avoids cross-interference

CachingStable prefix + session affinity routing — maximize prefix cache hit rate

Representative Technical Decisions

Logit Masking over Dynamic Tool Removal

Mutating the tools array breaks KV cache; use decode-time logit masking to constrain the action space instead

todo.md Attention Engineering

Update todo.md every iteration to keep the global plan near the recent end of context — counteracts “lost in the middle”

Error Retention Policy

Keep failure records so the model gets implicit belief updates — error-recovery behavior emerges naturally

→ Context Engineering · Prefix Caching · Context ManagementManus Blog (2025)

Manus

简介

Manus 是一家 AI agent 公司，开发了同名的通用自主 agent 产品。在上下文工程（context engineering）领域，Manus 团队以公开分享生产级 agent 架构经验著称，其博客文章《Context Engineering for AI Agents: Lessons from Building Manus》（作者：Yichao “Peak” Ji）成为行业重要参考。

对 Context Engineering 的贡献

Manus 经历了四次架构重建（自称”Stochastic Graduate Descent”），发展出一套生产验证的 context engineering 五维框架：

维度	核心策略
Context Offloading	文件系统作为无限外部 memory，零 token 成本
Context Reduction	仅限可恢复压缩（保留路径/URL 指针）
Context Retrieval	文件搜索工具作为结构化检索层
Context Isolation	子 agent 在独立 context window 中运行
Context Caching	Stable prefix engineering + session affinity 路由

核心经济洞察：agent loop 的 prefill/decode 比例约 100:1，prefix cache 命中率（Claude Sonnet：命中 $0.30/MTok vs 未命中$ 3.00/MTok）是首要生产成本杠杆。

代表性技术决策

Logit Masking 替代动态工具移除：在 agent loop 中修改 tools 数组会破坏 KV cache。Manus 保持完整工具集不变，通过解码阶段的 logit masking 约束动作空间。

todo.md 的注意力工程：通过在每次迭代中更新 todo.md 文件，将全局计划维持在 context 近期位置，对抗”中间迷失”（lost-in-the-middle）退化。

错误保留策略：刻意保留失败记录和错误信息，为模型提供隐式 belief update，使错误恢复行为自然涌现。

Manus

Manus

Manus

简介

对 Context Engineering 的贡献

代表性技术决策

相关来源

References