Glean 拾遗
中 EN
Submit a link
← All issues
#001 Latest 5/25–5/31 Published May 31

Beyond Prompts: The Rise of Harness Engineering and System-Level AI Thinking

This week's Glean traces a clear arc: we're moving decisively from the era of 'how to prompt better' into the depths of building engineering systems around AI. With Claude Code's Dynamic Workflows, ClickHouse's official Agent Skills, and Claude Managed Agents gaining cross-session 'Dreaming' and self-assessment, a clear signal emerges—the easy wins from single-point efficiency are plateauing. The real divide now lies in who can integrate context engineering, harness engineering, and organizational restructuring into repeatable, governable productivity systems. From Karpathy-inspired CLAUDE.md philosophies to ByteDance's data-backed 'context as moat' thesis, we've curated the week's most essential practices, reflections, and engineering blueprints. This edition is for every builder who refuses to be merely a user of AI.

38 picks 6 sections ~9 hr
Section 01

Paradigm Shifts: From Prompt to Context to Harness

5 / 38
x.com · 12 min
01

Context Engineering Is Replacing Prompt Engineering. Here's How It Works上下文工程正在替代提示工程：五层框架与实践指南

The author argues that prompt engineering is giving way to 'context engineering'—building the environment of information (identity, knowledge, memory, tools, processes) that enables a model to produce consistent results with minimal prompting. A five-layer framework is detailed, with practical steps for Claude users: set custom instructions, upload knowledge files, actively craft memory, connect MCP tools, and encode processes as Skills. The piece is opinionated and lacks empirical evidence but offers actionable guidance for those heavily using Claude.

AIFrameworkLLM
x.com · 3 min
02

Prompt → Context → Harness: The Three Paradigms of AI Engineering从 Prompt 到 Context 再到 Harness：AI 工程的三次范式转移

AI engineering has undergone three paradigm shifts: from Prompt Engineering (2023–2024) to Context Engineering (2025), and now to Harness Engineering (early 2026). Harness Engineering combines evaluation feedback loops, architectural constraints, and memory governance. Anthropic’s evaluator agent turned a 20‑minute useless artifact into a 6‑hour complete game; OpenAI built a million‑line system with zero human‑written code in five months, enforcing architectural boundaries via CI/linters. Two academic papers fill the memory layer: (S)AGE uses Byzantine‑fault‑tolerant Proof of Experience consensus to double agent calibration accuracy; a longitudinal study shows that 3 lines of prompt plus memory matches 200 lines of expert prompt in performance, yet only the memory group improves over time. Essential for engineers building multi‑agent systems.

AgentsAILLM
www.infoq.cn · 18 min
03

From OTel to Rotel: 4x Throughput Increase in PB-Scale Tracing从 OTel 到 Rotel：每秒处理量提升 4 倍的 PB 级追踪系统

This article benchmarks OpenTelemetry data planes for writing trace spans to ClickHouse. On the same 8‑core host, Rotel achieves 3.7 million spans/sec (462k spans/core/sec), a >4× improvement over the OTel Collector. Gains come from three optimizations: binary encoding of JSON columns in RowBinary, moving deserialization to a shared thread pool to avoid tokio blocking and glibc allocator lock contention, and enabling fast LZ4 compression. The test also exposes silent data loss in the OTel Collector under backpressure. For engineers scaling large telemetry pipelines.

DatabasePerformance
x.com · 20 min
04

how to build a production grade ai agent构建生产级AI智能体的十条工程原则

Over 40% of agentic AI projects fail, not because of models, but due to poor risk controls, architecture, and business value. This article presents ten engineering principles: threat modeling, strictly typed tool contracts, least-privilege execution, compact context engineering, governed retrieval, deterministic orchestration, separated memory, reliability mechanics, full observability, and continuous governance. Each principle provides concrete implementation details and real-world numbers (e.g., prompt injection appears in 73% of deployments), guiding teams to build secure, scalable production-grade agents.

AgentsAILLM
tw93.fun · 31 min
05

Agent Unveiled: Principles, Architecture, and Engineering PracticesAgent 架构、工程实践与落地：从原则到 OpenClaw

This article systematically examines the underlying architecture and engineering practices of agent systems. Starting from a stable agent loop, it contrasts workflows with agents, explains five control patterns, and emphasizes that the harness (evaluation baselines, execution boundaries, feedback, and fallbacks) often matters more than the model itself. It details context engineering via layered management and three compression strategies to prevent context rot, ACI‑oriented tool design, a four‑type memory system with consolidation, long‑task state externalization across sessions, protocol‑based multi‑agent coordination, eval frameworks (Pass@k and Pass^k), and event‑driven observability. Finally, it shows how these principles are implemented in OpenClaw, providing a practical reference for engineers building robust agents.

AgentsFrameworkLLM
Section 02

Claude Ascendant: From Chat Window to Autonomous Engineering Agent

12 / 38
x.com · 10 min
06

How to Actually Use Claude. 18 steps that unlock 100% of its potentialClaude 实战指南：18 个步骤解锁全部潜力

This guide provides 18 actionable steps to fully leverage Claude. It covers setting up Projects and Custom Instructions for persistent context, shifting your mindset to treat Claude as a thinking partner rather than a search engine, and using advanced techniques like style cloning, Extended Thinking, and token-saving prompts. Ready-to-use templates are included for Feynman-style learning, travel planning, expense analysis, and business idea stress-testing. A key insight: simply specifying output length can cut token usage by 40-60%. Aimed at users who want to go beyond basic Q&A and make Claude work for them.

AILLM
x.com · 11 min
07

Claude Can Do All of This. Most People Have No Idea.Claude 隐藏功能全指南：17 个你不知道的用法

This guide covers 17 hidden Claude features: persistent memory via Projects, interactive app building with Artifacts, step-by-step reasoning in Adaptive Thinking, long-term user profiling with Memory, role-based prompts, a browser agent (Claude in Chrome), desktop file-system access (Cowork), scheduled tasks, installable skills, CLAUDE.md project rules, terminal coding with Claude Code, visual design with Claude Design, and 90% cost reduction through Prompt Caching. Each includes where to find it and a ready-to-use prompt.

AI
claude.com · 12 min
08

Using Claude Code: The unreasonable effectiveness of HTMLHTML 的超常效力：用 Claude Code 摆脱 Markdown

Thariq Shihipar argues for using HTML instead of Markdown when working with Claude Code. HTML can represent tables, SVG, designs, and interactions—far denser information than Markdown. HTML docs are more readable, shareable, and can include interactive elements. Claude Code can pull context from codebases, Slack, git history to generate rich HTML reports, prototypes, and review interfaces. Concrete use cases cover planning, code review, design, reporting, and custom editing tools, with reusable prompt examples. For developers seeking to make Claude Code outputs more engaging and actionable.

AILLM
claude.com · 19 min
09

Claude Code in Large Codebases: Best Practices and Getting StartedClaude Code 在大代码库中的工作方式：最佳实践与入门指南

This article covers how Claude Code navigates large codebases using agentic search instead of RAG indexing, avoiding stale index issues but requiring good context configuration. It details the 'harness' ecosystem around the model—CLAUDE.md, Hooks, Skills, Plugins, MCP servers, LSP integration, and subagents—and presents three configuration patterns from successful deployments: making the codebase navigable, maintaining CLAUDE.md as models evolve, and assigning ownership for rollout. A practical guide for teams adopting Claude Code at scale.

AgentsAI
x.com · 10 min
10

From Scratch: Build Automated Claude Code Workflows with Hooks从 0 开始：用 Hooks 打造自动化 Claude Code 工作流

A tutorial on using Claude Code Hooks to automate shell commands at lifecycle events, replacing unreliable prompt instructions. Covers 5 key events (PostToolUse, PreToolUse, etc.), 3 hook types (command/prompt/agent), and config file structure. Provides 5 ready-to-use examples: desktop notification, auto-formatting, file protection, context recovery after compaction, and commit message linting. Exit code 2 blocks dangerous actions and feeds stderr back to Claude. For developers seeking reliable Claude Code workflows.

AgentsAI
tw93.fun · 20 min
11

Deconstructing Claude Code: Architecture, Governance, and Engineering Practices拆解 Claude Code：六层架构、治理与工程实践

Based on six months of intensive use of Claude Code, the author breaks down its functionality into six layers (CLAUDE.md/rules/memory, Tools/MCP, Skills, Hooks, Subagents, Verifiers) and provides design principles, anti-patterns, and configuration examples for each. The article focuses on context engineering (token cost structure, layered loading strategy, compaction pitfalls), tool design, Hooks for mandatory enforcement, Subagents for context isolation, prompt caching, and verification loops. Ideal for engineers wanting to move from ad‑hoc chat to a disciplined agent engineering workflow.

AgentsAI
claude.com · 6 min
12

Dreaming, Outcomes, and Multiagent Orchestration in Claude Managed AgentsClaude Managed Agents 推出「梦境」、成果评估与多智能体编排

Anthropic launches Dreaming in research preview for Claude Managed Agents: a scheduled process that reviews past sessions and memory to extract patterns, enabling agents to self-improve. Outcomes let developers define rubrics with a separate grader for self-correcting work; internal benchmarks show +10pp task success, +8.4% on docx, +10.1% on pptx. Multiagent orchestration allows a lead agent to decompose tasks to specialist subagents running in parallel with shared filesystem and traceability. Case studies include Harvey (6x completion rate improvement), Netflix (parallel log analysis), Spiral (writing quality via outcomes), and Wisedocs (50% faster document reviews). For engineers building autonomous AI agent systems.

AgentsLLM
x.com · 10 min
13

How to Make Claude Code Fix Its Own Mistakes Automatically (Exact Setup You Can Copy)让 Claude Code 自动修复自身错误的完整配置

This article provides a complete, copy-paste-ready setup for making Claude Code automatically catch, fix, and learn from its own mistakes. It covers a self-growing CLAUDE.md for project rules, PostToolUse hooks for auto-formatting and type-checking, Stop hooks for running tests on completion, PreToolUse hooks for blocking dangerous operations, and cross-session memory. The included settings.json config reduces back-and-forth from 45 minutes to 10 unattended minutes per feature. Audience: engineers using Claude Code or AI coding assistants.

AgentsAI
claude.com · 59 min
14

Best Practices for Computer and Browser Use with ClaudeClaude Computer Use 工程落地指南：缩放、缓存与思考

Official best practices guide for integrating Claude's computer and browser use capabilities, covering screenshot scaling, click accuracy, cache breakpoints, context management with rolling buffer and server-side compaction, prompt injection defenses, thinking effort tuning, and experimental features like batch tools and the advisor tool. Based on internal testing with Claude 4.6 and Opus 4.7, includes concrete code and performance data.

AgentsLLM
x.com · 16 min
15

How to Use Claude at 100% — Most People Never Get Past 10%Claude 100% 使用指南：大部分人只用到了 10%

This guide reveals 17 hidden features of Claude that most users never use, including Projects, Artifacts, Extended Thinking, Memory, Claude in Chrome, Cowork, Scheduled Tasks, Skills, CLAUDE.md, Claude Code, Claude Design, and Prompt Caching. Each feature comes with setup instructions and ready-to-use prompts. Perfect for anyone wanting to turn Claude from a simple chatbot into a full productivity system.

AgentsAILLM
x.com · 9 min
16

How I set up Claude to actually get work done我把Claude调教成工作系统：25个实用配置

Most people use Claude as a one-off Q&A, losing context each time. The author shares a systematic setup: personal instructions, projects, reference files, a context file, connected tools like email and calendar, templates, and repeatable workflows. 25 concrete steps transform Claude from a chat window into a reusable work environment. Suitable for technical workers frustrated with inconsistent AI responses.

AgentsLLM
x.com · 15 min
17

CLAUDE.md Guide: 21 Instructions to Lock In Preferences and Context用 CLAUDE.md 为 Claude 装上永久记忆：21 条配置指令指南

Most Claude users don't know about CLAUDE.md — a plain-text file placed in a project folder that Claude reads automatically at the start of every session, permanently setting your preferences, context, and behavioral rules. This guide provides 21 concrete instructions across five parts: communication style (no filler, admit uncertainty, match length to task), behavior (ask before big changes, only change what was requested, summarize changes), user context (background, project, writing voice), memory & continuity (log decisions in MEMORY.md, session summaries, track failures), and developer-specific rules including Andrej Karpathy's 4 golden rules (don't assume, simplest solution, don't touch unrelated code, flag uncertainty), which reportedly boosted coding accuracy from 65% to 94%. For anyone who wants to stop repeating themselves and get more consistent, on-brand output from Claude.

AILLM
Section 03

Developer Practices: Rules, Skills, and Automated Workflows

7 / 38
x.com · 14 min
18

ByteDance TRAE AI Coding Manuals: Context Engineering as Moat字节TRAE AI编程手册精读：上下文是护城河

A distilled summary of ByteDance TRAE team's 20 internal AI coding practice manuals. The core argument is that the bottleneck in AI coding efficiency is not model capability but context engineering. The article details six methodologies: Context Engineering, Skills, Spec Coding, Rules, MCP, and Agentic Coding, backed by experimental data (e.g., 32 real bug fixes: 100% success with Skills vs 59% without). Suitable for frontline developers, tech leads, and engineering managers.

AgentsAILLM
x.com · 13 min
19

5 Agent Skill Design Patterns Every ADK Developer Should Know5 种 ADK 开发者必备的 Agent Skill 设计模式

With SKILL.md format standardised across 30+ agent tools, the real challenge is content design. This article distills five recurring patterns from ecosystem-wide practices: Tool Wrapper (on-demand library context), Generator (template fill-in for consistent output), Reviewer (checklist scoring by severity), Inversion (agent-led interview before acting), and Pipeline (strict multi-step with gate conditions). Each pattern includes working ADK code, helping developers build reliable agents.

AgentsAIFramework
vercel.com · 6 min
20

Introducing React Best Practices: A Structured Repo for AI AgentsVercel 发布 React 最佳实践仓库，面向 AI 编程代理优化

Vercel distills 10+ years of React and Next.js optimization into a structured repo with 40+ rules across 8 categories, ordered by impact from eliminating waterfalls to JavaScript micro-optimizations. Each rule includes code samples and impact ratings, and compiles into an AGENTS.md document consumable by AI coding agents.

AIPerformanceReact
x.com · 5 min
21

10x Your Claude Skills with Karpathy's Autoresearch Method用 Karpathy 的 autoresearch 方法，将你的 Claude Skills 效果提升10倍

This article shows how to automatically improve Claude Skills using Karpathy's autoresearch method. It works by giving an agent a yes/no checklist and letting it iteratively test, tweak, and keep only beneficial changes. The author improved a landing page copy skill from 56% to 92% pass rate in four rounds, with visible changelogs. The method applies to any measurable task—define a checklist and let the agent run. Includes download links and practical examples for engineers building AI workflows.

AgentsAI
x.com · 2 min
22

Andrej Karpathy wrote something that every Claude Code user has felt b卡帕西三句话，说破每个Claude Code用户的痛点

Andrej Karpathy's three observations about LLM behavior—making silent assumptions, overcomplicating code, and performing careless side effects—inspired a single CLAUDE.md file with four principles: think before coding, prioritize simplicity, make surgical changes, and execute goal-driven. Each principle directly addresses a specific pain point. The file is ready to drop into any project to guide AI coding assistants toward more disciplined output. For every Claude Code user who has experienced these issues but struggled to articulate them.

AILLM
danielabaron.me · 12 min
23

CSS Refactoring with an AI Safety Net用AI安全网重构CSS：零视觉变化的七阶段计划

The author refactored a tangled CSS codebase into a clean architecture using Claude Code and Playwright, ensuring zero visual changes across seven phases. A Playwright script captured 9 app states, and after each phase, Claude compared screenshots to baseline, catching regressions like a line-height shift. The result: layered CSS with modern reset, unified button classes, and CSS variables. The post details state enumeration, script writing, and AI-driven diffing, and discusses trade-offs with dedicated tools. Essential reading for front-end developers tackling legacy CSS.

AIFramework
claude.com · 7 min
24

Introducing dynamic workflows in Claude CodeClaude Code 推出动态工作流：端到端跑完大型任务，自动并行调度与校验

Claude Code now supports dynamic workflows, enabling parallel orchestration of tens to hundreds of subagents within a single session for large-scale engineering tasks. It handles end-to-end jobs like codebase-wide bug hunts, migrations across hundreds of files, and security audits. Workflows dynamically plan, fan out, cross-validate, and converge results. Example: Bun was ported from Zig to Rust in 11 days, producing ~750k lines with 99.8% test pass rate. Workflows show plans before execution, can resume after interruption, but consume significantly more tokens. Available in research preview for Max, Team, Enterprise users.

AgentsAI
Section 04

Reimagining Organization, Mindset, and Collaboration

8 / 38
x.com · 19 min
25

Why Your “AI-First” Strategy Is Probably Wrong为什么你的“AI 优先”战略可能错了

The CTO of an agent platform shares their journey of rebuilding the entire engineering workflow around AI: 99% of production code is written by AI, shipping features within a day. The article critiques the superficial “AI-assisted” approach and introduces “harness engineering,” detailing their tech stack, self-healing feedback loop, and the new engineer roles of Architect and Operator. Real-world results include 3–8 deployments per day. Valuable for teams and CTOs seeking genuine AI integration.

AgentsAI
www.akashbajwa.co · 12 min
26

The Future Of Software Engineering with AnthropicAnthropic 圆桌：软件工程的未来

A summary of a roundtable on the future of software engineering, featuring leaders from Stripe, NVIDIA, Microsoft, and others. Key insights: closed-loop development creates compounding gains; test-first is the new default; human code review is fading; comments are written for AI readability; long-horizon tasks remain unsolved; developer tooling is being displaced first; hiring values experimentation over raw skill; human-authored context files help, agent-authored ones can hurt. Candid trade-offs and real-world practices are shared.

AgentsAILLM
x.com · 12 min
27

How Coding Agents Are Reshaping Engineering, Product and Design编码代理如何重塑工程、产品与设计

Coding agents are fundamentally reshaping the EPD (Engineering, Product, and Design) collaboration model. With the cost of implementation plummeting, the traditional PRD→mockup→code waterfall is dead, replaced by a review-centric process where prototypes are rapidly generated and then scrutinized. Generalists who wield coding agents gain unprecedented leverage; system thinking and product sense become essential for everyone. The bar for specialization rises, and roles converge into either builders or reviewers. Ultimately, anyone with a deep grasp of both product and technology can thrive, blurring traditional role boundaries.

AgentsAI
www.bassimeledath.com · 22 min
28

The 8 Levels of Agentic EngineeringAgentic 工程的八个层级：从自动补全到自治团队

Bassim Eledath maps the progression of AI-assisted coding into 8 levels, from tab-complete and AI IDEs to context engineering, compounding engineering, MCPs & skills, harness engineering with automated feedback loops, background agents, and autonomous agent teams. Each level builds on the previous, with practical insights on closing the gap between model capability and practice. He argues that plan mode is fading, multi-model dispatching yields better results, and true autonomous teams are still experimental. The piece serves as a roadmap for engineers looking to leverage AI more effectively.

AgentsAILLM
articles.zimetic.com · 8 min
29

Beyond the Coding Assistant — A New Series超越编码助手：企业级 AI 辅助软件工程系列开篇

This free series examines AI-assisted software engineering at enterprise scale. While individual coding speed has skyrocketed, many teams have not seen delivery improve—some have even slowed down. The author argues that current AI coding assistants optimize a single role, but software is shipped by teams with many non-coding roles. The next frontier is lifecycle orchestration, not better code generation. The series is structured in four parts, publishing three times a week with no paywall. It is aimed at engineering leaders, architects, and developers interested in AI engineering.

AgentsAI
x.com · 2 min
30

Organizational Structure for AI-First in the Harness EraHarness 时代 AI-First 的组织架构

A podcast interview with Creao's founders explores Harness Engineering—building self-healing, self-improving systems around LLMs. True AI-First companies restructure around AI as the primary producer: development cycles shrink from weeks to a day, product managers are dismantled, and cross-team alignment is automated. Junior engineers adapt faster than seniors; the future rewards architecture + product + marketing generalists. The 'Agent Economy' means content may be produced for AI consumers. A 25-person team rebuilt their architecture in two weeks. Full transcript available.

AgentsAIFramework
www.koshyjohn.com · 11 min
31

A.I. Should Elevate Your Thinking, Not Replace It工程思维的分水岭：你是在用 AI 提升层次，还是外包思考？

Software engineering is splitting into two groups: those who use AI to remove drudgery and invest in higher-level thinking, and those who outsource their reasoning to AI, simulating competence without building it. This 'outsourced thinking' is a new failure mode that erodes judgment over time. The real value of engineers lies in framing problems, making tradeoffs, and creating clarity—skills AI cannot own. Early-career engineers are especially at risk of skipping essential skill formation. Leadership must learn to differentiate polished output from genuine technical depth. The article argues that organizational health depends on recognizing this divide.

AI
spin.atomicobject.com · 7 min
32

Your Best Prompt Is a Well-Defined User Story你的最佳提示就是一份定义清晰的用户故事

In the age of agentic development, user story quality directly impacts AI output. The article argues teams should invest more time in breaking down stories and writing clear acceptance criteria rather than just estimating story points. A well-defined story includes three parts: Context, Acceptance Criteria, and Technical Hypothesis. Story point estimation is valuable only when forecasting or surfacing team misalignment is needed; otherwise it can be skipped. A good story acts as a good prompt, accelerating development cycles. Relevant for engineering teams using agile/Scrum.

AgentsLLM
Section 05

Infrastructure in Action: ClickHouse and a New Stance on Observability

4 / 38
clickhouse.com · 3 min
33

Introducing ClickHouse Agent SkillsClickHouse 发布 Agent Skills：28 条规则让 AI 助手学会 ClickHouse

ClickHouse has released official Agent Skills: an open-source set of 28 prioritized best-practice rules covering schema design, query optimization, and data ingestion, packaged using Anthropic's Agent Skills specification. Users can add them locally with `npx skills add clickhouse/agent-skills`. AI agents (e.g., Claude Code) automatically invoke these rules when appropriate, helping avoid common pitfalls like wrong ORDER BY, non-scalable JOINs, or missing materialized views. The Apache 2.0-licensed repo welcomes community contributions.

AgentsAIDatabase
www.infoq.cn · 15 min
34

ClickHouse 10 Best PracticesClickHouse十大最佳实践：从主键到JOIN的深度优化指南

A ClickHouse solution architect shares 10 field-tested best practices derived from customer engagements, covering schema design, data types, partitioning, skipping indexes, JSON type, data ingestion, materialized views, system tables, ReplacingMergeTree, and JOIN optimization. Benchmarks on a 150M-row Amazon reviews dataset quantify the impact: proper ORDER BY reduces rows scanned by 347×, unnecessary partitioning slows queries by 46×, correct data types cut storage by 12% and double query speed, skipping indexes reduce scans by 80%, and dictionary lookups beat regular JOINs by nearly 3×. The article emphasizes understanding ClickHouse internals to achieve orders-of-magnitude improvements without hardware changes.

DatabasePerformance
clickhouse.com · 15 min
35

ClickStack Observability: MCP Server, AI Notebooks, and ClickStack CloudClickHouse 可观测性三连发布：MCP Server、AI Notebooks 与 ClickStack 云服务

At Open House, ClickHouse announced three major observability updates: ClickStack Cloud (serverless, managed, private preview), AI Notebooks (beta), and an open-source ClickStack MCP server. AI Notebooks replace linear chat with persistent, branchable investigation workspaces, exposing every query and step. The MCP server provides semantic investigative tools to external agents; internal benchmarks show 25% fewer tool calls, 2.5× consistency improvement, and 20% higher evaluation scores vs. raw SQL MCP. The server also supports bi‑directional orchestration: agents can create dashboards and persist results. The design philosophy is “bring your own agents,” with SQL as an escape hatch when pre‑built tools fall short. The post includes setup instructions and a demo. For infrastructure/SRE engineers evaluating ClickHouse-based observability.

AgentsAIDatabase
blog.cloudflare.com · 18 min
36

Project Glasswing: What Mythos Showed UsGlasswing 项目：Mythos 的安全漏洞挖掘实战与启示

Cloudflare tested Anthropic's Mythos Preview on 50+ internal repos under Project Glasswing. The model excels at chaining low-severity bugs into working exploits and generating PoCs, making validation actionable. Real-world use revealed inconsistent model refusals and signal-to-noise challenges; a generic coding agent proved ineffective. Cloudflare built an eight-stage harness (Recon, Hunt, Validate, Gapfill, Dedupe, Trace, Feedback, Report) using parallel narrow tasks and adversarial review to improve quality. The post argues that beyond faster patching, defenses must limit exploit reachability from the architecture layer.

AgentsInfraLLM
Section 06

More

2 / 38
x.com · 19 min
37

Claude Code Dynamic Workflows: A New Primitive That Moves Orchestration Into CodeClaude Code Dynamic Workflows：把编排逻辑搬进代码的新原语

Anthropic introduces Dynamic Workflows, a primitive that turns task orchestration into JavaScript scripts executed by a deterministic runtime. The script manages loops, branching, and intermediate results, so only the final answer enters the main Claude context—solving the bottleneck of context overflow and attention dilution when coordinating hundreds of parallel tasks. A deep dive into architecture, primitives, and execution model is paired with a real-world Bun-to-Rust migration (11 days, 750K lines, 99.8% test pass) and a personal 133-session analysis. Comparisons with n8n/Coze/Dify show Workflow's advantage: Turing-complete code offers more expressiveness than visual DAGs, and the orchestration can be generated on‑the‑fly by a model. It shines for codebase audits, large migrations, and adversarial verification but comes with high token costs and current preview limits. Target audience: engineers tackling massive, automated coding tasks.

AgentsAIFramework