Recommendations 建议
路线图这就来了
30 Matching Annotations
- Jun 2026
-
-
-
The result is "almost-right code … the most expensive kind of wrong" because reviewers must reconstruct the missing system context for every change
几乎正确的代码是代价最高昂的错误, 丢失了上下文
-
The inner-loop/outer-loop framing
内外环模型 https://www.reddit.com/r/AI_Agents/comments/1n1zjwd/the_outer_loop_vs_the_inner_loop_of_agents_a/
The inner loop is the control cycle of a single agent that hat gets some work (human or otherwise) and tries to complete it with the assistance of an LLM. The inner loop of an agent is directed by the task it gets, the tools it exposes to the LLM, its system prompt and optionally some state to checkpoint work during the loop. In this inner loop, a developer is responsible for idempotency, compensating actions (if certain tools fails, what should happen to previous operations), and other business logic concerns that helps them build a great user experience. This is where workflow engines like Temporal excel, so we leaned on them rather than reinventing the wheel. 内环是单个智能体的控制周期，它接收某些工作（来自人类或其他来源），并尝试在 LLM 的协助下完成。智能体的内环由其接收的任务、向 LLM 开放的工具、系统提示词（system prompt）以及可选的用于在循环期间记录工作检查点的状态所引导。在内环中，开发者负责幂等性、补偿操作（如果某些工具失败，之前的操作该如何处理）以及其他有助于构建卓越用户体验的业务逻辑关注点。这正是像 Temporal 这样的工作流引擎所擅长的领域，因此我们选择依赖它们，而不是重新造轮子。
The outer loop is the control loop to route and coordinate work between agents. Here dependencies are coarse grained, where planning and orchestration are more compact and terse. The key shift is in granularity: from fine-grained task execution inside an agent to higher-level coordination across agents. We realized this problem looks more like proxying than full-blown workflow orchestration. This is where next generation proxy infrastructure like Arch excel, so we leaned on that. 外环是用于在智能体之间路由和协调工作的控制循环。在这里，依赖关系是粗粒度的，规划和编排更加紧凑和简洁。关键的转变在于粒度：从智能体内部的细粒度任务执行，转向跨智能体的高层级协调。我们意识到这个问题看起来更像是一种代理（proxying）行为，而不是完整的工作流编排。这正是像 Arch 这样的下一代代理基础设施所擅长的领域，因此我们选择了依赖它。
-
AI-generated UIs match target designs at 83–89% fidelity, with failures "localized and non-catastrophic" — fine-grained styling/spacing rather than structural errors.
细粒度的样式问题而非结构性错误 -> 视觉回归测试集成到 CI -> 是否真的有用?
-
Callstack (a React Native consultancy) is publishing optimization guides specifically because "AI coding agents are writing more code — including React Native apps" and need codified best practices for the JS-native bridge, FPS, and time-to-interactive (the two core metrics they map skills to).
https://www.callstack.com/blog/announcing-react-native-best-practices-for-ai-agents
-
'feeling' of speed."
"错觉" 这个用词很好, 感觉很快但是实际受限于机制
-
adapt AI to the process,
PR/CI 搞一个自动化 (走个人额度或者我的
-
It helps teams with strong process foundations and harms those without.
架构的重要性
-
AI accelerates the "inner loop" (coding) while the "outer loop" (review, security, compliance, deployment)
这个生命周期的架构很适合写进 PPT
-
the dominant failure mode is "almost-right" code
90% 成功, 剩下失败, 也不可能发布
-
-
-
具体的工具细节（如 Cursor 的规则格式、Opus 4.8 / GPT-5.5 等模型版本）很快就会过时。
需要在 PPT 中强调
-
CodeRabbit's Dec 2025 analysis found AI co-authored code carried ~1.7× more "major" issues. Treat technical debt and security regressions as default risks, not edge cases.
技术债和安全性是默认风险
-
"AI for AI's sake"
这不就是
-
Stage 2 — Build the harness on one team (weeks 4–12).
正在做的
https://www.reddit.com/r/ExperiencedDevs/comments/1rbhxsv/anyone_else_trimming_down_aigenerated/ - AI 很容易以极高的效率，生成一堆完全符合 Lint 规范、却在架构上极其短视或臃肿的“垃圾工程” - 架构设计充满了“隐性知识”与折衷（Trade-offs）
人类定骨架，AI 填血肉: 明确定义好项目结构、API 边界、规范文件 https://martinfowler.com/articles/harness-engineering.html
当绿地项目逐渐演变为一个“边界清晰、规范确立”的成熟 codebase 时，才是从零构建并启动 Harness 自动化控制网格的最佳时机 https://www.reddit.com/r/opencodeCLI/comments/1trz990/what_do_you_use_to_start_a_greenfield_project/
-
LLM's "vibes" in a way "better suited for a weekend project, rather than for customer-facing applications of a publicly listed company" (CodeRabbit).
适合 sideproject 而非面向客户的生产级应用
-
A practical "form factor" summary: the minimal credible AI-native workflow deliverable is a Git repo containing AGENTS.md (router) + docs/ (system of record) + .claude/skills/*/SKILL.md (procedures) + Spec Kit gates + test/CI/observability hooks + a settings/permissions file. Open-source scaffolders (e.g., harness-creator skills) can generate this skeleton in minutes.
一个 ai native workflow 脚手架可以 apply 到所有项目
-
5. How to build an AI-native workflow from zero
PPT 部分, 可以介绍目前我们有的, 正在做的, 限制有哪些, 下一步做什么
-
知识图谱，以及将工作分解为“代理就绪型任务（agent-ready tasks）
确认哪些能给 agent 做, 哪些不行
-
The convergent architectural components
这个架构可以画到 PPT: rule | knowledge (packaged) | scaffolding (spec-driven) | harness (feedback loops)
-
machine-readable artifacts
机器可读的资产
-
harness: "the system around the model. It decides how work gets split, which subagents spawn, what tools each one gets, how their output is verified... and when the job is actually done" (Anthropic).
harness 的真正定义: 一个治理框架
-
Humans steer. Agents execute.
老生常谈
-
four-level progression
四个步骤, 可以放到 PPT 里面, 我们处在 2->3
-
describe intent, accept AI output without reviewing the code
不经审查
-
Andrej Karpathy (OpenAI co-founder, former Tesla AI director) coined "vibe coding" on February 2, 2025, in a post on X: "There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."
vibecoding 起源 https://x.com/karpathy/status/1886192184808149383
-
On the one-year anniversary (Sequoia AI Ascent 2026), Karpathy declared vibe coding obsolete for professional work and proposed "agentic engineering."
agentic engineering https://www.ibm.com/think/topics/agentic-engineering
-
"AI-native" is distinct from "AI-assisted/AI-augmented." IBM's definition: "AI-augmented systems rely on AI as a supporting tool, whereas AI-native systems are AI-driven at their core... if the AI were to be removed, the product would not just cease to function as intended, it would cease to be useful at all."
由 AI 驱动, 移除 AI 则无法按预期运行
-
value comes from rearchitecting workflows, not from handing developers tools.
核心价值来源于工作流重构而非仅仅提供更强的工具
-
is not a prompt — it's a harness.
不是一个提示词, 而是一个系统框架 指令 + 路由 + 打包的知识 (skill) + 开发脚手架 + 反馈循环
-
a system where humans steer and agents execute, anchored by a reusable, version-controlled artifact (a "harness").
最后是一个 version-controlled artifact -> harness
-