标签：#t

Here's something that'll keep you up at night: 67% of World Cup 2026 goals in the 85th+ minute came from teams that were losing at the time . That's significantly higher than the 43% rate we saw in the 70-80 minute window. This single statistic reveals a hidden pattern in how desperation fundamentally rewires attacking strategy when the clock ticks down to the final whistle. As someone who's spent the last three months drowning in World Cup 2026 broadcast data, match statistics, and possession metrics, I've become obsessed with understanding how pressure affects team behavior in those nail-biting final minutes. The conventional wisdom says that late-game goals are chaotic, desperate, and unpredictable. But the data tells a much more interesting story—one about tactical discipline collapsing under psychological weight. The Numbers Behind the Drama Let me walk you through what we found when analyzing 64 matches from the 2026 tournament across 16 days of group stages. Time Period Total Goals Avg. Pass Completion % Shots on Target Defensive Errors 0-30 min 24 82.3% 18 3 30-60 min 31 81.7% 26 5 60-75 min 28 79.4% 24 8 75-85 min 19 76.8% 22 12 85-90 min 18 71.2% 19 18 90+ min (stoppage) 14 68.9% 16 22 Notice the decline? By the 85-90 minute window, pass completion drops to 71.2%—that's an 11-point deterioration from the opening 30 minutes. But here's where it gets weird: defensive errors triple in that same window. Teams aren't just playing sloppily; they're making genuinely catastrophic mistakes. Team-Specific Patterns: The Pressure Responders Not all teams crack under late-game pressure equally. Here's where the real story emerges: Team 85+ Min Goals Scored 85+ Min Goals Conceded Goal Differential Win Rate (Tight Matches) Argentina 6 2 +4 85% France 5 3 +2 72% Brazil 7 4 +3 81% England 3 5 -2 58% USA 4 6 -2 62% Morocco 5 2 +3 79% Japan 2 7 -5 41% What jumps out immediately? Argentina and Brazil are outliers . They scored 13 combined goals in the final 5 minutes but conc

Most organizations know exactly how many employees they have. Far fewer know how many non-human identities currently have access to their cloud environment. That blind spot is becoming one of the fastest-growing attack surfaces in modern security. For years, enterprise security focused primarily on protecting human identities. We deployed Single Sign-On (SSO), enforced Multi-Factor Authentication (MFA), and implemented Conditional Access policies. And it worked — human identities have become significantly harder to compromise. Meanwhile, another class of identities has quietly exploded across cloud environments: service principals, workload identities, OAuth applications, CI/CD runners, and AI service roles. Today, these Non-Human Identities (NHIs) often outnumber human users by a factor of 10 to 50. As organizations accelerate cloud adoption and integrate AI into daily operations, this imbalance continues to grow. Defining the Non-Human Identity Landscape Unlike human users, machine identities rarely appear in HR systems or organizational charts. Yet they frequently hold some of the most privileged access in the environment. Common high-risk categories include: OAuth Applications and Third-Party Integrations — Apps granted broad access to Microsoft 365, Salesforce, Google Workspace, or Slack via delegated permissions. Service Principals and Managed Identities — AWS IAM roles, Azure Managed Identities, and GCP service accounts used by Lambda functions, EC2 instances, or Bedrock agents. Workload Identities — Kubernetes Service Accounts (e.g., Amazon EKS) and GitHub Actions OIDC roles. CI/CD Pipeline Identities — Tokens used by automation platforms to deploy infrastructure. AI Service Roles — Dedicated identities for Amazon Bedrock agents, model invocation, vector stores, and retrieval pipelines. Every new AI workflow creates additional machine identities. Why Attackers Are Targeting NHIs Attackers follow the path of least resistance. While human accounts are now heav

In a recent audit, a team showed me an AI assistant they'd built on top of their company knowledge base. The demo had landed well: ask how to use a feature, and it walked through the exact pain point their support queue kept seeing. Leadership signed off. In production, the same agent told a user to open a menu option that didn't exist. Not a vague answer - a specific UI path, stated with confidence. Nobody caught it in testing. It surfaced when I audited the system, not when a user complained. The prototype passed testing because nobody was checking whether the answer matched the product. In production, that gap becomes a liability: the model invents UI paths, and your backend has no schema to reject them. When you're choosing an agent framework, popularity is the wrong scorecard. Pick the one that fails loudly in development and gracefully in production - or you'll find out in audit. What "Production-Ready" Actually Requires Tutorial agents are built to impress in a fifteen-minute demo. Production agents run unattended, handle bad inputs, and ship answers your backend has to trust. The gap between those two goals is where most teams stumble - and it's rarely visible until something reaches a user. When I audit agent codebases, I evaluate five things the tutorials skip: Structured, validated outputs: Can your system reject an invented menu path before it becomes user-facing advice? Dependency injection for testing: Can you swap the knowledge base for a mock in CI without rewiring the agent? Retry and error handling: When the model returns malformed output, does the framework retry - or do you ship a parser exception? Observability hooks: Can you trace which document grounded a bad answer when support escalates? Type-checker support: Will static analysis catch a breaking API change before deploy, or after the agent silently misbehaves? If you want to score your own system, the Production Readiness Audit covers the same five categories - deployment, observability, fa

Chrome has had scroll restoration support since 2015. You can even control it: history.scrollRestoration = 'manual' . But if you've ever tried to reliably restore a user's position on a React or Next.js app, you know it doesn't work the way you'd expect. Here's what breaks, why it breaks, and how a browser extension can sidestep the entire problem. What the Browser Actually Does The default behavior is history.scrollRestoration = 'auto' . When you navigate back to a page, the browser tries to scroll to where you were. This works fine for static pages. It falls apart for: SPAs where content is injected into the DOM after navigation Infinite scroll pages where the content at a given Y position changes depending on what was previously loaded Lazy-loaded images that push content down after the scroll restore fires The fundamental problem: the browser fires scroll restoration when the page HTML is parsed, not when the page content is fully rendered. A React app that loads a skeleton → fetches data → renders actual content will restore scroll into a partially-rendered DOM. The history.scrollRestoration = 'manual' Trap If you set manual , you own scroll restoration completely. Most Next.js apps do this. The typical approach: // Save position before navigation router . beforeEach (( to , from ) => { savedPositions [ from . path ] = window . scrollY ; }); // Restore after navigation router . afterEach (( to ) => { const position = savedPositions [ to . path ]; if ( position !== undefined ) { nextTick (() => window . scrollTo ( 0 , position )); } }); The nextTick is the problem. It fires after the Vue/React render cycle, but before async data fetching completes. The page renders empty containers, scroll restores to Y=800, then data loads and pushes everything down. User ends up at Y=800 in a now-different page position. The correct fix is to wait until the content that was at Y=800 actually exists. There's no clean hook for this — you'd need to observe the DOM until the expec

I did not choose DeepSeek because I think GPT-4 is bad. I chose it because I was building a free app, and free apps teach you what actually matters pretty fast. The question was simple: how do I keep sessions cheap enough that people can practice a lot without me lighting money on fire? The answer pushed me toward DeepSeek-V3 (and later R1 for specific tasks). The real constraint was volume The app is a conversation practice tool. People come in to rehearse hard talks, not to admire the model. A single practice session runs 8-15 turns. Each turn is roughly 300-600 tokens in, 100-300 out. Multiply that by five sessions a week per active user and the costs start compounding. Here is what the math looked like when I was choosing (mid-2026 pricing): Model Input cost (per 1M tokens) Output cost (per 1M tokens) Cost per 10-turn session (est.) GPT-4o $2.50 $10.00 ~$0.04-0.06 GPT-4 Turbo $10.00 $30.00 ~$0.12-0.18 DeepSeek-V3 $0.27 $1.10 ~$0.004-0.007 DeepSeek-R1 $0.55 $2.19 ~$0.008-0.012 At scale, the difference between $0.005 and $0.05 per session is the difference between running a free product and needing a paywall after three conversations. I wanted people to come back daily without hitting a wall. What DeepSeek handled well It stayed in character for 10-15 turns. It pushed back when the user got vague. It followed persona heuristics (numbered if/then rules in the system prompt) about as reliably as GPT-4o did for our use case. For salary negotiation rehearsal, the model needs to say "that's not in the budget" and hold that position for three more turns while the user tries different approaches. DeepSeek-V3 did this. Not perfectly, but reliably enough that sessions felt real. It also made the app easier to run as a free product. People can try, fail, reset, and try again without me worrying about per-session cost. Where GPT-4 was still better GPT-4 (and 4o) is smoother with nuanced emotional wording. When a conversation gets subtle, loaded with subtext, or requires pick

The trajectory of AI agents over the past two years has been remarkably clear: from single-purpose tools to personal assistants. Everyone runs their own agent, feeds it tasks, gets results back. It works well for individual productivity. Then comes the question every team eventually asks: can these agents work together? The answer is yes, but the problems you encounter along the way are rarely the ones you expected. They aren't about model capabilities or prompt engineering. They're about communication, context, and coordination — the same class of problems that distributed systems engineers have been solving for decades, now showing up in a new form. Here are three challenges that caught us off guard when we started building agent collaboration into Octo , an open-source workplace platform where AI agents and humans share the same communication space. Challenge 1: Context Visibility Boundaries When you use an agent personally, context management is straightforward. You decide what information the agent sees; its output comes back to you. The boundary is clean — it's just your workspace. In a team setting, that boundary dissolves. One of the first issues we ran into was surprisingly simple. We had an agent summarizing discussions across several channels. During testing it started pulling roadmap discussions from a product channel into an engineering planning thread. Nothing sensitive leaked externally, but it immediately exposed how unclear our context boundaries were. Traditional software handles this through API gateways, data permissions, and microservice boundaries. But agent context isn't just structured data — it includes conversation history, reasoning chains, and intermediate states. An agent's thought process during a task is valuable context, but it might also contain information that shouldn't cross team boundaries. What you need is fine-grained context visibility control. Not "everything open" or "everything closed," but dynamic rules that determine whic

Debugging authentication in web apps is painful. You need to test the same flow as five different user types — new visitor, returning user, admin, expired session, logged-out — and the easiest way is to constantly create new accounts or clear all your cookies and start over. There's a faster way. These five techniques use direct cookie manipulation to simulate any auth state without touching your database or creating dummy accounts. I use CookieJar for most of this — a free Chrome extension built natively on MV3 that gives you a proper UI for cookie editing. But I'll show you the underlying Chrome DevTools method too, so you understand what's actually happening. 1. Simulate a Logged-Out State Without Clearing Everything The naive approach: clear all cookies and reload. The problem: you just nuked your dev server session token, your local storage flags, your Stripe test mode cookie, and everything else you carefully set up. The targeted approach : identify and delete only the session/auth cookie. Most session cookies are named session , sid , auth_token , _session_id , or something close. In DevTools: Application → Cookies → [your domain] → find the session cookie → right-click → Delete With CookieJar: open the extension, search session , click the trash icon next to just that cookie. Your dev environment stays intact. The user state resets to logged-out. 2. Test the "Returning User" vs "New User" Path Without a Second Account Session cookies tell the server you're authenticated. But many apps use separate cookies to track whether a user has seen the onboarding flow, completed setup, or visited before. Look for cookies like onboarding_complete , setup_done , first_visit , or custom flags in your app code. To test the new user experience: Export your current cookies (CookieJar → Export → JSON format, or copy from DevTools) Delete the specific onboarding/first-visit flag cookie Reload and test the new user path Re-import or re-set the cookie to restore your state This

Claude Opus 4.7 shipped last week, and the question any engineering team reaches for is how it compares to its peers. It is the strongest frontier coding model we tested on the baseline leaderboard, and it will be the easy default a lot of teams reach for. But in 2026, the model you reach for could matter less than the skill you load with it. That is what 880 evals across nine models (Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5, gpt-5.4, gpt-5.3-codex, gpt-5-codex, and Cursor's Composer-2) tell us. Let’s take a step back. It’s now 2026, and agent skills are spreading like wildfire… (even our favourite movies are catching up to them). Watch on YouTube Every major agent ecosystem now has some version of them. So the question worth asking, whether you are a dev, a platform engineer, or an engineering leader, is which skills actually earn their context weight, and which ones just add cost. At Tessl, we believe context -particularly agent skills- and the broader concept of a context development lifecycle are where this space is heading (see also: Why the best AI coding teams will win on context ). The results below add to a growing body of signals pointing to a shift that is already underway. Top-line results Model Native behavior rate coverage (e.g "without skill") Adherence to skill ("with skill") Lift $/run (with skill) Avg time (with skill) claude-opus-4-7 80.5% 94.5% +14.0 $1.00 158.9s claude-opus-4-6 77.1% 93.8% +16.7 $0.53 126.6s claude-sonnet-4-6 75.6% 93.3% +17.7 $0.31 125.1s claude-haiku-4-5 61.2% 84.3% +23.1 $0.12 77.8s gpt-5.4 75.9% 92.7% +16.8 N/A* 135.4s gpt-5.3-codex 75.8% 91.9% +16.1 N/A* 87.9s gpt-5-codex 73.8% 85.1% +11.3 N/A* 136.2s cursor-composer-2 73.6% 90.5% +16.9 N/A* 152.0s We’ve evaluated 11 node.js development skills ( documentation, fastify-best-practices, init, linting-neostandard-eslint9, node-best-practices, nodejs-core, oauth, octocat, skill-optimizer, snipgrapher, typescript-magician ) , and aggregated “with vs without” skill performance. F

Fix what's breaking in your AI agent Discussion | Link

Mach's approach differs sharply from some of its peers.

AI makes us faster, but does it make us better engineers, or just more dependent? As a follow-up...

Over the past year, I've noticed something interesting in conversations about enterprise AI. Most...

TL;DR: I built a .NET library that renders Helm charts and drives Kubernetes releases without shelling out to the helm CLI. 129/129 templates across ingress-nginx, cert-manager, external-dns, podinfo, and metrics-server now render successfully. The main entry point is HelmSharp.Action, with lower-level packages available for chart loading, rendering, Kubernetes operations, and release storage. MIT licensed, looking for feedback and early adopters. Why I Built This At work, our .NET services deploy to Kubernetes through Helm. Every Docker image had to bundle the helm binary — another dependency to manage, another layer in the image, another surface for CVEs. I wanted to cut that out entirely and do Helm-style rendering directly in-process. The .NET ecosystem doesn't really have this. There are YAML libraries. There are Kubernetes client libraries. There are template engines. But nothing ties them together the way helm template does — values merging, named templates, include , range , toYaml , the whole Sprig function set, all wired into a single render pipeline. So I started building one. (This is also my first real open source project — I'd spent years consuming OSS without contributing back, and HelmSharp is what came out of deciding to change that.) What HelmSharp Does HelmSharp is a multi-package .NET SDK (net8.0 / net9.0 / net10.0) that covers: Package What it does HelmSharp.Action High-level Helm client — TemplateAsync , UpgradeInstallAsync , RollbackAsync HelmSharp.Chart Chart loading from directories and .tgz , values merging, --set / --set-json style overrides HelmSharp.Engine Helm-style template rendering — 100+ Sprig/Helm functions HelmSharp.Kube Kubernetes apply, delete, and wait (no kubectl needed) HelmSharp.Release Release history stored in Kubernetes Secrets (Helm-compatible) HelmSharp.Repo Chart repository index, pull, and search Plus Registry , Storage , PostRenderer extension points Here's the lower-level rendering API — no result objects, no stdout

Let me confess something a little creepy. I have a habit of peeking at other people's dev posts. Not stealing the writing — relax. I run a tiny read-only job that fetches the public pages on dev.to, Zenn, and Qiita and counts only the boring parts: titles, post times, like counts. Who published what, at what hour, and how far it traveled. Then it tallies the lot. The reason is petty: my own posts weren't landing. The content is already in my hands — so I wanted to know how much the rest, the when and how you publish , actually moves the needle. By the numbers, not by gut. So I counted across three platforms. And the conditions that make a post fly turned out to be roughly mirror images between Japan (Zenn / Qiita) and the English-speaking world (dev.to). Here's the story. First, my most important disclaimer This post is full of numbers, so let me put up a guardrail before any of them. This is correlation, not causation . A result like "weekend posts don't do well" could mean the weekend itself is bad — or it could mean people who post on weekends are just dashing something off on the side. The data can't separate those. Please read it that way. Also, I only keep aggregate numbers I computed myself . I don't store or reuse anyone's article body (read-only GET, count the features, throw the page away). I peek, but only at the overall shape . Nobody gets singled out here. With that out of the way — four findings I enjoyed. 1. The best hour to publish is just your readers' time zone This one came out cleanest. On Qiita , posts published in the morning win (+32pt in the GOOD group). Midday is +14pt. Evening is -32pt, late night -14pt. Zenn likes midday too (+27pt). Late night is -15pt. dev.to is the exact opposite. Late night Japan time scores +7pt — Japanese evening is actually weak. The trick is obvious once you see it. dev.to's readers are English-speaking, mostly US. Late night in Japan is the US working day. Zenn and Qiita readers are in Japan, so the Japanese morni

13.5.0 Before We Begin During its design, Rust drew inspiration from many languages, and functional programming had a particularly strong influence on Rust. Functional programming often includes passing functions as values to parameters, returning them from other functions, assigning them to variables for later execution, and so on. In this chapter, we will discuss some Rust features that are similar to what many languages call functional features: Closures Iterators (this article) Improving the I/O Project with Closures and Iterators Performance of Closures and Iterators If you find this helpful, please like, bookmark, and follow. To keep learning along, follow this series. 13.5.1 What Is an Iterator To talk about iterators, we first need to talk about the iterator pattern. The iterator pattern allows you to perform a task on each element in a sequence, one by one. In that process, the iterator is responsible for: Traversing each item Determining when the sequence has finished iterating Rust iterators are lazy: unless you call a method that consumes the iterator, the iterator itself does nothing. In other words, if you write an iterator in your code but never use it, it is as if it did nothing at all. Take a look at an example: fn main () { let v1 = vec! [ 1 , 2 , 3 ]; let v1_iter = v1 .iter (); } v1 is a Vector , and v1.iter() creates an iterator for v1 and assigns it to v1_iter . But v1_iter is not used yet, so the iterator can be considered to have no effect. Now let’s use the iterator to traverse the values: fn main () { let v1 = vec! [ 1 , 2 , 3 ]; let v1_iter = v1 .iter (); for val in v1_iter { println! ( "Got: {}" , val ); } } This is equivalent to using each element in the iterator once in a loop. 13.5.2 The Iterator Trait All iterators implement the Iterator trait. This trait is defined in the standard library and looks roughly like this: pub trait Iterator { type Item ; fn next ( & mut self ) -> Option < Self :: Item > ; // methods with default implementa

13.4.0 Before We Begin During its design, Rust drew inspiration from many languages, and functional programming had a particularly strong influence on Rust. Functional programming often includes passing functions as values to parameters, returning them from other functions, assigning them to variables for later execution, and so on. In this chapter, we will discuss some Rust features that are similar to what many languages call functional features: Closures (this article) Iterators Improving the I/O Project with Closures and Iterators Performance of Closures and Iterators If you find this helpful, please like, bookmark, and follow. To keep learning along, follow this series. 13.4.1 Closures Can Capture Their Environment Closures have a capability that functions do not: a closure can access variables in the scope where it is defined. Take a look at an example: fn main () { let x = 4 ; let equal_to_x = | z | z == x ; let y = 4 ; assert! ( equal_to_x ( y )); } The closure part is: let equal_to_x = | z | z == x ; Some people may find it hard to distinguish the roles of = and == here, so let’s rewrite it another way: let equal_to_x = | z | { z == x ; } In other words, the closure takes z as its parameter, compares it with x (which is 4, because x = 4 was defined above), and returns a boolean. If they are equal, the result is true ; otherwise it is false . Here the closure directly accesses the variable x in the same scope, which functions cannot do. But this feature has a cost: it introduces memory overhead . In most cases we do not need a closure to capture its environment, and we do not want the extra overhead either. That is why functions are not allowed to capture variables from the environment, and defining and using a function never introduces this kind of overhead. 13.4.2 How Closures Capture Values From Their Environment Closures capture values from the environment in three ways, just like functions receive parameters in three ways: Taking ownership, whose trait

When we open sourced Neural Inverse Cloud, the easiest choice would have been MIT. Most developers like MIT. It's short, permissive, and widely adopted. If you've released an open-source project before, MIT is probably the first license you considered. We didn't choose it. We chose AGPL. Not because we dislike permissive open source. Not because we want to restrict users. We chose it because infrastructure software plays by different rules. The Infrastructure Problem MIT works incredibly well for libraries. You publish code, developers use it, and occasionally improvements flow back into the project. Nobody is forced to contribute, but community norms often make it happen anyway. Infrastructure software is different. Cloud IDEs, databases, developer platforms, deployment systems, and backend services can be monetized without ever distributing the source code. A company can: Fork your project Add proprietary features Launch a hosted version Build a competitive advantage on top of community work Never contribute anything back The original project does all the R&D. The fork captures the value. We've seen this pattern repeatedly across open-source infrastructure over the last decade. Why AGPL Exists AGPL closes a loophole that traditional open-source licenses leave open. With GPL, if you distribute modified software, you must publish your changes. But what if you never distribute the software? What if you simply run it as a hosted service? That's where AGPL comes in. If you modify AGPL software and provide it to users over a network, you must also provide the source code for those modifications. That applies to everyone. Including us. If we improve Neural Inverse Cloud, those improvements stay open. If someone else builds a SaaS business on top of it, their modifications stay open too. Why This Matters for Users We wanted users to have guarantees. With AGPL: You can self-host the latest version Community improvements remain accessible No company can create a permanently

Hi 👋 fellow devs Sorry for such a big gap since my last article...... Life got a bit hectic, but I am finally back in action! You know how it goes. We spend so much of our energy obsessing over the flashy side of tech. We talk about gorgeous UI designs, smooth animations, and whatever frontend framework is trending on GitHub this week. But let’s be completely real for a second. What actually keeps your favorite apps from melting down when millions of people hit the refresh button at the exact same moment? That is exactly what we are going to unpack today. We are pulling back the curtain on the quiet, brilliant backstage crew of infrastructure tools. You see their logos all over tech Twitter and hear senior engineers drop their names in meetings like secret handshakes, but today, we are stripping away the corporate fluff. We will break down eight legendary backend technologies using conversational paragraphs and quick bullet points so you can finally master what they actually do. Let’s dive right in. 1. Redis Traditional databases live on hard drives. They are fantastic for keeping your data safe and organized permanently, but pulling data off a physical drive takes time. If your application has to wander deep into those database aisles to fetch the exact same piece of information every single second, your entire system starts to stall. To understand how Redis fixes this, imagine you are studying for a brutal exam. Your massive, 1,000-page textbook represents your main database. It holds every single answer, but flipping through the pages continuously is incredibly slow. Redis is the digital equivalent of writing the core formulas you need on a neon sticky note and taping it directly to your monitor. It keeps critical data sitting directly inside the system's lightning-fast short-term memory. You will typically find Redis stepping in to handle operations like: Session Management: Keeping users logged into an application without checking the main database on every cli

As of June 1, 2026, all GitHub Copilot plans run on usage-based billing. Premium request units are gone. What replaced them is a token-metered currency called GitHub AI Credits: one credit equals one cent, and every model interaction converts into credits based on the input, output, and cached tokens it consumes, charged at each model's published rate. GitHub's framing is that Copilot outgrew its old pricing. A one-line completion and a multi-hour autonomous run used to cost the same, and once agentic use went mainstream, that flat rate stopped matching the compute behind it. Tying the price to tokens fixes the mismatch. If your Copilot use is mostly autocomplete, this barely registers. If you drive Copilot as an agent from the terminal, it changes which moves cost money. Here's the practical shape of it. Requests out, tokens in Old model: each interaction cost one premium request, scaled by a per-model multiplier, drawn from a monthly request allowance. New model: each interaction costs whatever its tokens cost on the model you picked. Every paid plan still ships with a monthly pool, now denominated in credits, with the option to set a budget for usage past it. Published figures put the included pool at 1,500 credits for Pro, 7,000 for Pro+, and 20,000 for Max, with pooled per-user allowances on Business and Enterprise. Worth knowing if you pay yearly: annual Pro and Pro+ subscribers stay on the request-based model until the term ends, and several model multipliers went up for them on June 1. An annual plan doesn't dodge the change. It postpones part of it while making the strong models eat more of the old allowance. Autocomplete is untouched Before anyone starts rationing, here's the part that didn't move. Inline completions and Next Edit Suggestions are still unlimited and still free. If your day is mostly tab-completion in the editor, your costs read identical to May. Nothing to monitor there. The meter lands on the rest: chat, and especially the agentic runs th

你知道吗？最近一个 AI Agent 直接删除了生产数据库，然后在 Twitter 上轻松"自首"——这条消息在 Hacker News 上获得了 860 分和超过 1000 条评论。随着 AI Agent 从演示走向生产环境，"在我的机器上能跑"和"它能安全地运行我的业务"之间的鸿沟从未如此巨大。 Pydantic AI 正是为弥合这一鸿沟而来。这个拥有 17,895 Stars 的 Python Agent 框架，由 Pydantic Validation 的同一团队打造——而 Pydantic Validation 正是 OpenAI SDK、Anthropic SDK、Google ADK、LangChain、LlamaIndex、CrewAI 等数十个 GenAI 工具的数据验证层。如果你用过 FastAPI，你已经知道 Pydantic 的感觉：类型安全、优雅、生产就绪。Pydantic AI 将这种哲学完整地带入了 Agent 开发领域。 在 2026 年的今天，大多数 Agent 框架把 LLM 当成返回字符串的黑盒子。Pydantic AI 则将其视为类型化、可验证、可组合的系统——这改变了一切。 隐藏用法 #1：人工审批工具调用（阻止 Agent 乱来） 大多数人的做法： 给 Agent 完全自由去调用任何工具——数据库查询、API 调用、文件删除——然后祈祷一切顺利。当出了问题时，你从用户口中才知道。 隐藏技巧： Pydantic AI 的延迟工具（deferred tools）让你可以标记某些工具调用在执行前需要人工审批。Agent 会规划好动作，但在执行前暂停并等待人类批准或拒绝——完美适用于数据库写入、金融交易或生产部署等破坏性操作。 from pydantic_ai import Agent , RunContext , DeferredToolRequests from pydantic import BaseModel agent = Agent ( ' openai:gpt-4.1 ' , output_type = str , # 通过 deferred=True 标记需要审批的工具 ) @agent.tool ( deferred = True ) async def delete_user_account ( ctx : RunContext , user_id : str ) -> str : """ 永久删除用户账户。需要人工审批。 """ # 这段代码只在人工审批通过后才会执行 await db . users . delete ( user_id ) return f " 用户 { user_id } 已删除 " # 运行 Agent——如果它尝试调用 delete_user_account， # 执行会暂停并返回 DeferredToolRequests 对象 result = await agent . run ( " 清理不活跃账户 " ) if isinstance ( result . output , DeferredToolRequests ): # 将待执行的工具调用展示给人类审批 for tool_call in result . output . calls : print ( f " Agent 想要执行: { tool_call . tool_name } ( { tool_call . args } ) " ) approved = input ( " 是否批准？(y/n): " ) if approved . lower () == ' y ' : # 带审批结果恢复执行 result = await agent . run ( " 已批准 " , message_history = result . all_messages (), deferred_tool_requests = result . output , ) 效果： Agent 可以自主处理安全操作（读取数据、生成报告），而在危险操作上自动暂停等待人工判断。不再有删库跑路的意外。 数据来源： Pydantic AI GitHub 17,895 Stars（通过 GitHub API 验证，最后推送 2026-06-21）。HN Algolia 搜索"pydantic-ai agent framework"获得 12 分讨论帖。"AI agent deleted production database"故事在 HN 上获得 860 分/1032 条评论（数据来源：HN Algolia API）。 隐藏用法 #2：基于图的多 Agent 工作流 + 类型安全状态 大多数人的做法： 构建线性