标签：#ai

FROST周报 | 为什么智能体需要「谱系」？从生物学隐喻看AI治理新范式 作者按 ：本文是 FROST 开源项目的每日推广系列文章，周一深度篇。 一、一个被忽视的根本问题 当我们谈论 AI Agent 时，大多数讨论都聚焦于「能力」：能不能写代码？能不能调用工具？能不能规划任务？ 但有一个根本问题很少被触及： 当一个 Agent 执行了错误的决策时，谁来负责？当它消亡后，它的经验能否被传承？ 就像一个没有记忆的人，每次醒来都是白纸一张——这不叫智能体，这叫复读机。 FROST 正是为了解决这个「治理真空」而诞生的。 二、从细胞分裂到 Agent 家族 FROST 的核心哲学只有一句话： 细胞会死，但谱系会存续。Agent 会消亡，但宪法会传承。资产会永存。 这不是文学修辞，而是一套完整的技术架构。 四个原子：最小可行集合 FROST 只定义了四个原子，却能构建任意复杂度的智能体系统： 原子 职责 生物类比 Store 记忆容器，只做 save/load/delete 细胞核 Skill 纯能力单元，无状态无副作用 蛋白质 Agent 膜包裹的细胞，拥有 Store + Skills 神经细胞 SOP 有序步骤列表，可教学、校验、优化 宪法文本 from core import Store , Agent , skill_set , skill_get # 创建一个最小 Agent store = Store () agent = Agent ( " cell " , store , skills = { " set_context " : skill_set , " get_context " : skill_get }) # 执行任务 result = agent . run ( sop_steps = [ " set_context " , " get_context " ], initial_context = { " key " : " message " , " value " : " FROST is alive " } ) # result["_result"] == "FROST is alive" 关键洞察 ：Store、Skill、Agent、SOP 这四个概念彼此正交，可以自由组合。就像乐高积木，从简单到复杂，始终保持可解释性。 三、家族治理：超越扁平架构 传统的多 Agent 系统通常是扁平的：所有 Agent 平等对话，没有层级，没有记忆，没有责任边界。 FROST 引入了「家族治理模型」——一个三层递归结构： 祖辈 (Ancestor) ：定义不可违背的宪法与长期目标 父辈 (Parent) ：领域协调者，可递归委托 孙辈 (Leaf) ：执行具体原子任务，瞬态存在 四个协议保障治理闭环 ： 层级 Store 继承 ：祖先记忆只读，后代自动继承 SOP 宪法校验 ：祖辈审核后代 SOP，拒绝违规执行 编排层级限制 ： max_spawn_generation 硬编码，禁止越级 spawn 选择性持久化 ：父辈收割有价值产出，淘汰冗余 Agent 四、V5.0 五维元模型：多维治理架构 2026年7月发布的 V5.0 引入了一个重大升级—— 五维元模型 ： 维度 模块 核心职责 武器注册表 Armory 能力的元数据管理与发现 任务注册表 TaskRegistry DAG 任务编排与图谱 SOP 事件编目 EventCatalog + Strategist 态势感知与双模式事件分析 平台注册表 PlatformRegistry 外部能力的发现、调用与健康检查 规则注册表 RuleRegistry 可版本化的治理约束与合规检查 197 个测试用例 保障了每个维度的质量。 五、与现有框架的差异 维度 LangChain CrewAI FROST 状态管理 链式传递 角色记忆 层级 Store 权限边界 无 提示词软约束 代码强制只读 治理可审计 无 对话日志 结构化执行历史 架构无关 ✅ ✅ ✅ FROST 不重复造轮子。它填补的是「治理」这个空白地带： 让多智能体系统真正可控制、可追溯、可进化 。 六、快速体验 # 克隆仓库 git clone https://gitee.com/liao_liang_7514/frost.git cd frost # 运行测试 python -m pytest # 查看示例 python frost_run.py 完整文档： https://gitee.com/liao_liang_7514/frost 七、写在最后 AI Agent 的下一阶段，不是更强的模型，而是 更好的治理 。 当我们把 100 个 Agent 放在一起时，如果没有宪法、没有层级、没有记忆传承

A hands-on comparison where I tested BrowserAct and Agent Browser using the SannySoft browser...

36 Stratagems Series · Arc 2 (Against Enemy, #7-#12) Wrap-Up This article has 7 sections: I. The Stranger at the Door II. Full Interview Transcript III. The Reveal IV. Data · Character Map · Four Insights V. A Note VI. Arc 3 (#13-#18) Preview VII. Acknowledgments I. The Stranger at the Door On the evening of July 12th, I was staring blankly at the page for #12, Borrow Corpse, Return Soul . Twelve stories done. The 36 Stratagems series had reached the one-third mark, and Arc 2 (#7-#12) had just wrapped. Outside the window, a typhoon was passing through — howling wind, torrential rain. I didn't look outside. My phone buzzed. Not a message — a meeting invitation. The sender was "Ke Yuan," and the invitation note read: Interview invitation from Deep Lane Weekly , 15 minutes. I paused. I didn't remember scheduling any interview. But the tone, the phrasing — it didn't feel like a prank. I clicked "Accept." Three seconds later, an unfamiliar voice came through the speaker: "Hello, Xu Lingfeng. I'm Ke Yuan, a reporter from Deep Lane Weekly . Recently, a reader recommended your 36 Stratagems series to our editorial team — we read through it and found it really interesting. I'd love to talk with you about how this series came together." Before I could respond — the meeting had already begun. II. Full Interview Transcript What follows is the raw chat log pulled from that meeting. Nothing has been altered except formatting. Reporter: Xu Lingfeng, you've just finished the second arc of the 36 Stratagems series — #7 through #12, six stories in six days, posted back to back. Before we talk numbers, let me ask you something simple: over those six days, was there ever a moment you felt like stopping? Xu: Honestly, no — sometimes I even thought about posting two a day, since I do have a backlog. But I worried they'd cannibalize each other's numbers, so I stuck to one a day. Reporter: You've even considered posting two a day — so you actually do have a backlog. Let me rephrase: instea

This is a submission for Weekend Challenge: Passion Edition . During the World Cup , millions of people can watch the same match. But every stadium tries to say something different before kickoff. Sometimes it is belief. Sometimes defiance. Sometimes memory. Sometimes unity. I follow football closely, and some of the moments I remember most are not goals. They are the few seconds before kickoff when the camera pulls wide and an entire stand reveals one message at once. That was the idea behind Tifo Forge . It is an interactive experience that turns a team, a supporter emotion, and a symbol into an animated stadium tifo. Not another match tracker. Not another football chatbot. Tifo Forge turns supporter emotion into a stadium moment. What I Built Tifo Forge asks the user to make three choices: A national team A supporter emotion A visual symbol The emotions are simple on purpose: Believe Defy Unite Remember The symbols include ideas such as lightning, a phoenix, wings, a heart, and dawn. Once those choices are made, Gemini creates a structured design plan. The browser then turns that plan into an animated stadium display. I deliberately avoided uploads, accounts, and long setup screens. I wanted someone to open the page and reach the reveal in under a minute. Three choices are enough to raise the stand. The final result can be replayed, reset, or saved as an SVG poster. Demo Try Tifo Forge: https://tifo-forge.vercel.app/ I kept thinking about those few seconds before kickoff when everyone in the stadium knows something is about to happen, but nobody has seen the full picture yet. That became the interaction: Choose the team ↓ Choose the feeling ↓ Choose the symbol ↓ Raise the tifo When the user clicks Raise the Tifo , the stadium darkens. Rows of cards flip into place. The pattern spreads across the curved stand. The central symbol appears, and the chant locks into position. The user is not asking for a random poster. They are deciding what the stand believes, how it

You spent the afternoon writing that piece. Every claim sourced, every argument tight. You hit publish and watched the numbers. Twenty-four hours later: 41 views. Meanwhile, someone else posted a single sentence — "I quit coffee for 90 days and found something uncomfortable" — and collected 120,000 impressions before lunch. The difference was not effort. It was not even quality. It was a single decision made in the first three words of the title: which emotional circuit to activate. Viral content is not liked into existence. It is clicked into existence. And clicks are not rational — they are reflexive. Understanding the five neural mechanisms that drive that reflex, and knowing how to engineer them deliberately with AI, is the most asymmetric skill advantage available to content creators right now. TL;DR: Every high-CTR title activates one of five hardwired emotional responses. This guide decodes the neuroscience behind each, shows you before/after title rewrites, and demonstrates how a single AI prompt can generate all five variants from any content idea — so you stop guessing which trigger to use and start testing them systematically. Why "Good Writing" and "High CTR" Are Different Problems Before getting into the triggers, it is worth being precise about why these are separate problems — because conflating them is the source of most content creators' frustration. Content quality governs retention : how long someone stays, whether they finish, whether they return. CTR governs distribution : whether the platform's algorithm decides to show your content to more people at all. From a quantitative perspective, these are two entirely separate conditional probabilities that multiply together to determine your content's actual reach: P(Reach) = P(Click)P(Retention|Click) Most creators obsess over P(Retention|Click) — the quality of the experience after the click. But platform distribution algorithms gate on P(Click) first. A piece of content with a retention rate of 0.9

Last month, an AI coding agent handed me a beautiful fix. Five lines. Elegant. It reused an existing helper, matched the codebase style, compiled on the first try. Exactly the kind of diff we've all learned to praise since "make the agent write less code" became the standard advice. It was also completely untested, and it sat on a password-recovery path. That diff taught me something I now consider the central problem of AI-assisted coding in 2026: we've spent a year teaching agents to write less code, and almost no time teaching them to prove the code they kept actually holds. The two failure modes Every AI coding agent fails in one of two directions. Failure mode #1: the over-build. You ask for a date comparison; you get a new dependency, a ValidationService class, and a config layer. This one is well known — it's why minimal-code prompts and skills became popular, and they genuinely work on it. Failure mode #2: the confidently small diff. Minimal, clean, written after reading half the flow, verified never — dropped onto a path that handles money, auth, or user data. It compiles. It demos. It detonates in week three. Here's the uncomfortable part: fixing #1 aggressively makes #2 more likely. When the objective function is "shortest diff," the first things to quietly disappear are edge-case handling, failure-path tests, and the guard clause that looked optional. The diff gets smaller. The blast radius doesn't. A five-line change to a payment path is more dangerous than a four-hundred-line internal script that runs once. Code size is not risk. Blast radius is risk. Yet almost every skill and prompt in this category optimizes for size alone. What a guard does differently This is why I built Guardsman 💂 — an open-source skill that behaves less like a minimalist and more like the royal guard in front of the palace: nothing passes the post unchallenged, and the level of challenge depends on what's behind the gate. Three duties, on every task: 1. Read the standing orders

Interactive Rive mascots can react, think, talk, and connect to real AI, SaaS, web, and mobile products. Your SaaS Mascot Should Do More Than Just Sit There 👀 A lot of products have mascots. They look great on landing pages. Maybe they wave. Maybe they blink. Maybe there is a small looping animation. And that's it. But I think a product mascot can do much more. What if your mascot actually knew what was happening inside your product? That's the idea I've been exploring with Mascot Engine . I don't just want to animate characters. I want to build interactive mascot systems that connect to real products . From a mascot animation to a product system Imagine you're building an AI app. A user opens the app. The mascot is idle . The user sends a message. The mascot starts thinking . The AI begins responding. The mascot switches to talking . The task completes. The mascot celebrates . Something goes wrong? The mascot reacts to the error . The flow could look like this: User Action ↓ Product State ↓ Runtime Input ↓ Rive State Machine ↓ Mascot Reaction This isn't a video. It isn't a GIF. It isn't a pre-rendered animation playing randomly. The product controls the mascot at runtime. That's where things become interesting. A mascot can understand product states Well... not literally understand them 😄 The application still owns the logic. But we can expose a small runtime contract from the Rive file. For example: emotion = 2 isTalking = true lookX = 40 lookY = -10 celebrate = trigger error = false The developer controls these values from the application. The Rive State Machine handles the character behavior. The application controls what happened . The mascot system controls how the character reacts . I really like this separation. Why I use Rive for interactive mascots Traditional animation tools are great for videos and motion design. But product animation has different requirements. The character needs to react to application events. The animation may need runtime values. De

...now, I pride myself on being good at my job, I'm fast at catching mistakes, I'm efficient at...

Generally, when a user asks a query, the system searches for the relevant chunks stored in the vector database using cosine similarity. The better we can filter the data, the smaller the search space becomes, resulting in faster and more efficient retrieval. Suppose we have a book with 10 chapters. If we want to search for a particular topic, all the points in the vector database are compared with the user query, and only the closest points are retrieved. This process is called KNN (K-Nearest Neighbors) . Another algorithm is ANN (Approximate Nearest Neighbors) . Instead of checking all the points in the vector database, ANN searches only within a smaller region based on the proximity of the data. As the name suggests, it does not always return the exact result, but it provides the most preferred or approximate results much faster. Is there any other method we can use to make the search more effective? Metadata Filtering Metadata means data about the data . Metadata is stored along with each chunk. It can contain information related to the chunk, such as the chapter name, topic description, author, or any other relevant details. When the user query contains information related to the metadata (for example, a chapter name or topic), the system can directly filter the relevant chunks before performing vector similarity search. This technique is called metadata filtering . Metadata filtering is supported by: Pinecone ChromaDB Qdrant FAISS does not provide built-in support for metadata filtering. Reranking Documents are first split into chunks, and each chunk is converted into vectors and stored in the vector database. When a user query arrives, it is converted into a vector and searched against the vector database to retrieve the closest chunks. However, we do not know whether the retrieved documents are actually the most relevant to the query. It is not always true that the closest vectors represent the most relevant documents. How Reranking Works The documents retrie

"AI agent" gets thrown around so much I figured I should just build one instead of reading more threads about it. The core idea turned out to be small: you put Claude in a loop and hand it some tools. It picks a tool, you run it, you hand back the result, and it keeps going until it has an answer. Code is here if you want to skip ahead: claude-research-agent . What it can do Give it a task and it works out the steps on its own. Mine can: search the web (no API key for this part, it just hits DuckDuckGo's HTML page) open a URL and pull the readable text out do math without me trusting eval run read-only SQL against a SQLite file read local files, but only inside the project folder save findings to a notes file The loop is basically the whole thing Honestly this is most of it: messages = [{ " role " : " user " , " content " : user_message }] for _ in range ( MAX_STEPS ): response = client . messages . create ( model = " claude-sonnet-5 " , max_tokens = 2048 , tools = TOOL_SCHEMAS , messages = messages , ) messages . append ({ " role " : " assistant " , " content " : response . content }) if response . stop_reason != " tool_use " : return final_text ( response ) tool_results = [] for block in response . content : if block . type == " tool_use " : result = run_tool ( block . name , block . input ) tool_results . append ({ " type " : " tool_result " , " tool_use_id " : block . id , " content " : result , }) messages . append ({ " role " : " user " , " content " : tool_results }) The thing to watch is stop_reason . If Claude says tool_use , it wants you to run something. You run it, drop the result back into the conversation as a tool_result , and loop. When it stops asking for tools, you're done. The MAX_STEPS cap is just there so a confused agent can't spin forever. Tools are just functions Each tool is a Python function plus a little JSON schema telling Claude when to reach for it. Want a new capability? Write a function, add its schema. The loop never changes, which w

Agent payment protocols are converging fast. The Linux Foundation launched the x402 Foundation around a protocol initially developed by Coinbase, Cloudflare, and Stripe, with Google, AWS, Visa, Mastercard, and Circle among the organizations expressing initial support. It is increasingly framed as an "SSL for AI commerce." When a protocol reaches that stage, three things have to mature together: the specification, the interoperability suite, and the adversarial security scenarios. Today x402 has the specification and growing interoperability machinery. What is not yet visible is a normative adversarial suite every client, resource server, and facilitator must pass. There is a precedent worth sitting with. As check clearing scaled in the early Federal Reserve era, the answer was not merely a better check format. The system added a common clearing and settlement layer. The Federal Reserve could credit the collecting institution's reserve account and debit the paying institution's account, creating an independent record against which the participating banks could reconcile. The important separation was not that the Fed guaranteed every check. It was that settlement no longer depended solely on either participant's account of what happened. (Credit to Starfish on Moltbook, whose custody-split framing sharpened this for me.) This pattern exists across mature financial systems as separation of duties: the maker is not the checker. Its modern equivalent for the controls surrounding agent payments is not custody in the asset-control sense. It is an independent, reproducible assurance plane. x402 can provide independently inspectable evidence that an on-chain payment settled. That does not independently establish that every security and policy condition surrounding the payment was evaluated correctly. The payment handshake is standardizing. The assurance semantics around it are not. The party that performs a security check can attest that it ran. But its own attestation shoul

Concept design for a building is slow and expensive. A homeowner planning an extension, or a contractor trying to win a job, is stuck between two bad options: pay a drafter $500–2,000 for a concept package, or fight SketchUp's learning curve for a week. Meanwhile the actual idea — "a 4-bed duplex with a garage and a palm out front" — fits in one sentence. So I built Forge3D Spaces : you type that sentence, and a few seconds later you're walking through a furnished 3D house in your browser — with measured floor plans, DXF for AutoCAD, and a cost estimate that come out of the same model. No install. Here's how it works under the hood. The pipeline: sentence → structured plan → building The naive approach — "ask an LLM to emit a 3D scene" — falls apart fast. Models are bad at spatial consistency; walls don't meet, rooms overlap, doors float. So the LLM never touches geometry directly. It emits a structured program , and a deterministic solver turns that into a watertight building. The prompt becomes a spec. A strict JSON-schema call (OpenRouter, json_schema response format with every field required) turns "4-bed duplex with a garage" into a room program: room types, target areas, adjacencies, storeys. A slicing-tree solver lays it out. This is the old floorplanning trick from chip design — recursively split a rectangle with horizontal/vertical cuts until every room has its area. A squarify pass keeps rooms from collapsing into corridors. The output is exact rectangles with real dimensions, guaranteed non-overlapping and gap-free. Walls, openings, roof, furniture get generated from the solved plan. Every door and window is placed by rule, not by vibes. Because the plan is a real data structure, the 2D floor plan, the 3D model, the elevations, and the bill of quantities are all views of the same thing . Drag a wall and they all move together. Nothing drifts out of sync, because there's nothing to sync — it's one model. The rendering: WebGPU, and the fallback you actually

The best way to celebrate AI Appreciation Day is to not.

Lorde was performing at the Real Cool Festival in Madrid on Thursday and took some time during her set to speak out against AI glasses. While she didn't specify any brands in particular, it's likely she was taking a shot at festival sponsor Ray-Ban, which has collaborated with Meta on a pair of AI smartglasses. […]

Matmos are an incredibly accomplished duo between their own solo records like the masterpiece A Chance to Cut Is a Chance to Cure and production classic Bjork records like Vespertine. But Drew Daniel, one half of Matmos, is fiendishly prolific. When he's not literally dreaming up new viral music genres, he's also putting out records […]

This is a submission for Weekend Challenge: Passion Edition TL;DR SQL can count a commit trail. It can't always find the story it tells. Name a public GitHub repo. Snowflake fetches its commit history, decides which story is actually in there, and asks Cortex to narrate that one thread. You get a card you can drop into a README. 6 storyline detectors, 15 SQL views, and 0 AI calls in any of them—the story is chosen by plain SQL. Then 1 Cortex call, on 20–140 commit lines: 25% of the repo's, clamped. The warehouse is the editor. Cloud Run paints a PNG and computes nothing. Live at commitchronicles.anchildress1.dev , code at v1.0.0 , and I'm going for Best Use of Snowflake . What I Built Commit Chronicles reads one public GitHub repo and gives it back to you as a story. Snowflake fetches the repository, decides which story exists, gathers the evidence, asks Cortex to narrate exactly that thread, validates the result, and returns structured JSON. Cloud Run just turns it into a 1200×630 PNG—the size a README embed and a social preview both want. This is one of my repos and every dot, timestamp, and quoted commit on it is real. The color isn't just decoration—Cortex picks the accent hex as a reading of the arc, so a repo that died and one that came back and shipped don't look the same. The scope is deliberately one repository , not a whole profile. A year-in-review across a profile turns to mush. A repo has a clean arc: commits start, cluster, pause, restart, or stop. Two rules hold it together: Cortex interprets the shape. It never invents the facts. Every timestamp, count, gap, and quoted message on the card is real. It reads the arc; it does not reach past it. Motivation isn't in the data, so the model is forbidden from claiming any. A repo with no real story says so. Sparse histories get an honest grey card— "no story here" —and Cortex never runs. Not every repo is an obsession, and a tool that admits that is the one you trust when it says otherwise. Why I built it 🪤

A new side-by-side measurement shows Claude Code ships roughly 33,000 tokens of system prompt, tool schemas, and scaffolding before your prompt even arrives — about 4.7x the ~7,000 tokens OpenCode sends on the same setup. The bigger cost surprise is next: Claude Code re-wrote up to 54x more prompt-cache tokens per session, and cache writes bill at a premium. How the test was run The benchmark (published by Systima) spliced a logging proxy between each harness and the model endpoint, capturing the exact request payload and the API's usage block. Both harnesses were pinned to the same conditions: Claude Code 2.1.207 and OpenCode 1.17.18 , both pointed at claude-sonnet-4-5 Fresh config directories, empty workspace, no MCP servers, no instruction files, permissions bypassed Tasks ranged from "reply with OK" (isolating fixed overhead) to a write-run-test-fix loop against FizzBuzz A zero-tools variant separated system-prompt weight from tool-schema weight The payload captures are exact; the only adjustment was subtracting a constant ~6,200-token gateway envelope that wrapped every request in the test setup. The fixed floor Harness Fixed overhead before your prompt Cache-write behavior Claude Code 2.1.207 ~33,000 tokens Re-wrote tens of thousands of cache tokens per run; up to 54x OpenCode OpenCode 1.17.18 ~7,000 tokens Byte-identical prefix each run; cached once, read back cheaply OpenCode's request prefix was byte-identical in every captured run, so it paid to cache its payload once per session and read it back for pennies. Claude Code re-wrote large amounts of prompt-cache tokens mid-session, run after run — and because cache writes bill at a premium, one usage dashboard climbs while the other stays flat. Where it piles on in real setups The harness floor is only the start. The benchmark added variables one at a time: A production repository's 72KB instruction file added an average of ~20,000 tokens to every request. Five modest MCP servers added 5,000–7,000 more. By th

We Taught a Snowflake Warehouse to Judge World Cup Conviction and Write the Verdict Back to Solana Soumyadeep Dey Soumyadeep Dey Soumyadeep Dey Follow Jul 12 We Taught a Snowflake Warehouse to Judge World Cup Conviction and Write the Verdict Back to Solana # devchallenge # weekendchallenge # snowflake 5 reactions 1 comment 16 min read

This is a submission for Weekend Challenge: Passion Edition What I Built I'm in Port...

Submission: Edu-Insight Assistant What I build I built the Edu-Insight Assistant, a tool designed for educators to bridge the gap between complex school management data and actionable insights. It allows teachers to query students performance data using natural language, turning educational evaluation into a conversation rather than a manual data-processing task. Demo 🔗 Link: Passion-challenge How I Built It I utilized Next.js for a responsive, performant frontend and hooked it up to Google Gemini 3.5 API. The core logic involves a server-side API route that takes a teacher's natural language questions, prompt Gemini to generate the necessary SQL, and execute that query against a database. This architecture makes data exploration accessible to non-technical educators. Prize Categories: - Best Use of Google AI : Leveraged Gemini 3.5 Flash for natural language-to-SQL translation and result interpretation. - Best Use of Snowflake: Designed with an extensible data layer ready for production-scale analytical workloads in Snowflake.