AI 资讯

vibe citing: how KPMG used AI to write a report about AI and AI made them look like fools by t474-r0b07 There are companies that charge you to tell you how to use AI responsibly. KPMG is one of them. 250,000 employees. 138 countries. Decades advising governments and corporations on how to avoid costly mistakes. In October 2025 they published a report titled "Total Experience: Redefining Excellence in the Age of Agentic AI" . They wrote it with AI. The AI invented 88% of the sources. Nobody verified anything. They published it anyway. // what is "agentic AI" — because the title matters An agentic AI is not a chatbot. Not the assistant that answers your questions. It's a system that makes decisions and executes actions on its own, without a human approving each step. You give it an objective and it acts, corrects, moves forward. It's the product everyone in the tech sector was selling in 2025. KPMG was selling it too. That's why they needed a report proving their clients were already using it. Spoiler: they weren't. And the report invented it anyway. // the forensic analysis GPTZero — a company specialized in detecting AI-generated content — ran a full audit on the report. First: what is an AI hallucination, because the term is going to come up a lot. When a language model doesn't have the information you ask for, it doesn't say "I don't know." It generates a response that sounds correct. It invents with the same confidence it would use if it actually knew the truth. Perfect format. False content. No warning. That's a hallucination. Now the numbers from the KPMG report: TOTAL CITATIONS: 45 REAL CITATIONS: 5 INVENTED CITATIONS: 40 ACCURACY RATE: 11.1% 40 of 45 citations have invented titles, authors that don't exist, or sources that don't say what KPMG claimed they said. Half of the factual claims in the report are false or misattributed. A firm that charges for intellectual rigor published a document with 11% accuracy. // the organizations that read the report and sai

Few-shot learning with large language models is one of the most practical ways to steer model behavior without updating weights. By embedding task-specific examples directly into the prompt, developers can turn a general-purpose foundation model into a domain-specific classifier, parser, or reasoning engine. The technique relies on in-context learning, where the model infers patterns from exemplars rather than from gradient updates. Because it requires no training pipeline, few-shot prompting is ideal for rapid prototyping and production tasks where data volumes are too small for fine-tuning or where model weights must remain frozen. The Mechanics of In-Context Learning In-context learning is an emergent capability of transformer-based language models. During inference, the model attends to the full context window, using the provided examples as a dynamic prior. Each example adjusts the hidden-state activations for subsequent tokens, effectively conditioning the output distribution without any parameter change. Research suggests that the model locates latent task representations within its pretrained weight space and uses the few-shot examples to activate the appropriate subspace. The result is a flexible interface: change the examples, and the model adapts its behavior immediately. Zero-Shot, One-Shot, and Few-Shot Prompting These three patterns describe how much guidance you provide before the actual task input. Zero-shot: You describe the task in natural language with no examples. This works best for simple, well-known tasks that the model has seen frequently during pretraining. One-shot: You prepend a single example. This is often enough to communicate output format or tone. Few-shot: You prepend three to ten examples, sometimes more for complex schema extraction or multi-label classification. The marginal gain from each additional example typically diminishes, but for tasks with rigid output schemas, a larger set of exemplars can substantially improve consisten

Lately I keep reaching for SQLite where, before, I'd have reached for Postgres without thinking. It started with small services, then a bigger question: could a multi-tenant SaaS actually run on SQLite? And for AI agents specifically, isn't a local, embedded database the more natural home for their state? Turso is the version of this stack I've found most compelling so far, especially when paired with Cloudflare. I wish D1 would reach embedded-replica parity with Turso, and that AWS offered a managed SQLite-style service the way it offers RDS for Postgres. This isn't a "Postgres is over" argument. I still use Postgres more often than SQLite. And it isn't advice. It's just where my thinking has drifted recently — written down mostly so I can find out where it's wrong. Read it as one person's notes, not a recommendation. Where I've landed for now (and expect to keep revising): SQLite isn't replacing Postgres. For work state , it's increasingly my first reach, not my last. AI agents push this harder: their state is high-churn, local, and mostly private. The answer isn't all-local. It's a local workbench plus a central ledger . Why the old default existed For years, "where does the data live?" had one practical answer: a server, behind an API, in a shared Postgres. A lot of that wasn't architecture — it was the cheapest shape available. SQLite was already everywhere, but it lacked the operational layer that makes a database viable as SaaS infrastructure: networking, replication, managed backups, and a way to run many small databases without drowning in tooling. So centralizing was the path of least resistance, and a tenant_id column in shared Postgres became the reflex. What changed isn't SQLite. It's that the ecosystem grew the missing parts — and for a growing class of workloads, the thing doing the most frequent writing moved onto my own machine. The constraint that's lifting SQLite itself is, by design: Embedded, not networked — a library, nothing listens on a port.

We’ve all heard "it works on my machine," but when it comes to AI-driven features, that phrase is a recipe for disaster. You can have a perfectly tested agent today, but if you upgrade your base model or change your quantization strategy tomorrow, you might inadvertently kill your agent's reliability. You can’t afford to wait for production to find out your agent is hallucinating or failing its tool calls. This is why we built the headless QuantaMind CLI—to shift AI evaluation left into your CI/CD pipeline. By integrating custom eval JSON collections into your build process, you can now treat your AI agent like any other piece of code. If a model upgrade or a quantization tweak causes your agentic reliability to dip below your required threshold, your CI pipeline should block that merge. It’s not just about testing; it’s about enforcement. If you aren’t gating your deployments based on real, repeatable model performance, you aren’t shipping software—you’re shipping a guessing game.

AI agents are making software development faster. That is great. But there is a problem I do not think we are talking about enough: testing is not speeding up in the same way. In many teams, testing is still held together by spreadsheets, meeting notes, screenshots, chat messages, and the memory of a few experienced QA engineers. That worked when delivery was slower. It becomes fragile when one developer can use multiple agents to change code across several modules in a single afternoon. The bottleneck is no longer "can we write more test cases?" The bottleneck is: Can the team prove what was tested, why it was tested, what failed, what was fixed, and whether the release is safe? That is the problem I built testboat for. The Most Dangerous Sentence Before A Release The sentence I worry about most is not: We did not test this. At least that is honest. The dangerous sentence is: I think we tested this. That sentence usually means the team has test artifacts, but they are disconnected: requirements live in a doc test cases live in a spreadsheet automation scripts live somewhere in the repo execution results live in CI logs or chat bugs live in an issue tracker release reports are written manually before sign-off Each piece may be useful on its own. But when a Tech Lead asks, "Which requirements are not covered?" or a founder asks, "Can we release today?", the team has to reconstruct the answer manually. That is not a testing process. That is institutional memory under pressure. AI Makes This Gap Worse AI agents are very good at increasing throughput. They can: implement a feature faster refactor code faster generate UI faster write automation faster fix bugs faster But faster change creates more testing uncertainty. If an agent changes the authentication module, what should be rerun? If a test fails, is it a product bug, a flaky automation script, or an environment issue? If a developer says "fixed", has the failed test actually been rerun? If a release report says "ma

How to Call Windows Native APIs in Electron Calling Windows native APIs in an Electron app feels like wanting to see the ocean but only having a map. After some trial and error, I've found a few paths—writing this article serves as a record and a guide for others who might follow. Background When building Electron desktop applications, you inevitably need to interact with the operating system. On Windows, these requirements are quite common: Calling Windows Store APIs for in-app purchases Handling file system virtualization specific to Windows Store apps Obtaining system-level permissions and resources Interacting with Windows Runtime (WinRT) components Electron is fundamentally a Node.js environment, and Node.js doesn't natively provide direct access to Windows native APIs. A bridge is needed between the two. It's like trying to communicate with a friend who doesn't speak Chinese—you need a translator. Electron is written in JavaScript, Windows APIs in C/C++. The language barrier requires building a bridge. That's the harsh reality of the code world. About HagiCode The solutions shared in this article come from our practical experience with the HagiCode project. HagiCode Desktop needs to call Microsoft Store APIs to handle subscription purchases and license management, which is why we developed a set of technical solutions. After all, necessity drives innovation—that's a truth. Technical Solution Comparison When calling Windows native APIs in Electron, there are several mainstream approaches to choose from. Each has its applicable scenarios—like different tools in a toolbox, they work best when used in the right place, otherwise just add trouble. Solution Applicable Scenarios Pros Cons dynwinrt WinRT APIs (e.g., Store API) Type-safe, auto-generated bindings, modern JavaScript support Only supports WinRT APIs, requires Windows SDK Native Node.js Extensions High performance, any Windows API Complete control, optimal performance Requires C++ development skills, comple

I used to think a serious developer needed a serious IDE. Big project? Open PhpStorm. Design work? Open Photoshop. Need every refactor, every inspection, every plugin, every panel, every button? Load the heavy tool and wait for the machine to breathe again. But something changed. Not overnight, and not because those tools suddenly became bad. They are still powerful. The change is that AI started taking over the parts of the IDE I actually needed most. Today, I spend more time in VS Code and the terminal than in heavy IDEs. My machine feels lighter. My workflow feels less crowded. And honestly, I do not miss the old setup as much as I thought I would. The old IDE was a safety net For years, big IDEs won because they could see the whole project. They understood symbols, imports, frameworks, database models, refactors, formatting, inspections, and tests. A good IDE felt like a senior assistant sitting beside you, quietly warning you before you made a mess. That was valuable. It still is. But AI has started to move that intelligence out of the IDE shell. The useful part is no longer tied to one huge application. It can live in your editor, your terminal, your pull request, your CI pipeline, or even in a chat window with access to your codebase. When AI can read the files, reason about the bug, generate a test, run the test, inspect the failure, and propose a patch, the IDE becomes less like the brain of the workflow and more like one possible place to type. AI is becoming the environment The phrase "AI coding assistant" already feels too small. Autocomplete was the first version. The newer pattern is closer to an AI developer environment. You ask it to find the bug. It searches the repo. You ask it to explain a weird error. It follows the stack trace. You ask it to write a benchmark. It can create the benchmark file, run it, compare the result, and tell you what changed. You ask it to add tests. It can inspect the code path and generate cases you probably would have de

A pip-installable CLI tool that auto-centers off-center page number annotations created by macOS Preview, or batch-adds new ones — with smart content avoidance and landscape support. The Problem If you've ever used macOS Preview to add page numbers to a PDF (via the text annotation tool), you know the pain: numbers land wherever you drop them, never centered, and manually positioning dozens or hundreds of them is soul-crushing. Especially when the PDF has mixed portrait and landscape pages. I ran into this preparing a thesis — 200+ pages of final manuscript, page numbers visibly off-center on every single page. Editing each one by hand wasn't an option. The Solution pdf-pagenum is a single CLI command that reads a folder of PDFs and centers every page number annotation to the bottom of its page. It works by: Detecting FreeText annotations that look like page numbers Measuring body content boundaries on each page Repositioning the annotation to a clean, centered position below the content — with proper margins Preserving the original page dimensions (no resizing, ever) If your PDF has pages with no annotations at all, it can generate new page numbers from scratch in the correct position. Install pip install pdf-pagenum That's it. PyMuPDF and natsort come along as dependencies. Usage Fix Mode (default) Reposition existing page number annotations so they're centered at the bottom: pdf-pagenum ./scans/ ./output/ This is the mode you'll use 90% of the time — it takes whatever rough page numbers Preview gave you and snaps them to the mathematically correct center. Add Mode Generate brand-new page numbers on pages that lack them: # Number all pages starting from 1 pdf-pagenum ./scans/ ./output/ --add all # Number pages 3 through 7 only pdf-pagenum ./scans/ ./output/ --add 3-7 # Number specific pages, starting count from 10 pdf-pagenum ./scans/ ./output/ --add 1,3,5-7 --start 10 Ranges and comma-separated lists can be mixed freely. Start Offset The --start N flag works in b

Agentic frontend layer between Figma and Cursor & Claude Discussion | Link

See if Japanese AI search recommends your brand Discussion | Link

Ploy turns your website into your company's growth engine. Discussion | Link

Every tool call, every structured output, every agent decision travels as JSON. Here is the serialization knowledge that separates the amateur from the architect — now that the stakes have never been higher. A developer ships an AI agent on a Friday. In the demo it's flawless: the model reads a request, calls a tool, returns a clean answer the app renders perfectly. A week later, production dashboards are full of garbage. A date is showing up as raw text. A field that was definitely there is silently gone. Under one big payload, the whole server froze for two seconds. And here's the maddening part — nothing threw an error. The model returned JSON. The code parsed it. Everything "worked." The bug wasn't in the model, and it wasn't in the parser. It lived in the narrow gap between text and data — the place every JSON value has to cross twice. That gap is serialization , and in 2026 it has quietly become one of the most important things a JavaScript engineer can actually understand. Why now? Because the most important conversations in modern software aren't between humans anymore. They're between models and machines — an LLM deciding which tool to call, a server answering, an agent chaining ten steps together. And every one of those conversations happens in the same format: JSON. So let's open up the refinery and see how raw structure becomes a clean stream of bytes — and back again — without losing anything precious on the way. JSON is not a JavaScript object This is the misunderstanding that creates most JSON bugs, so it's worth saying plainly: JSON only looks like a JavaScript object. It isn't one. JSON is a transport format — flat, inert text meant to travel across a network or sit on a disk. A JavaScript object is a live structure in memory that your application can read, mutate, and call methods on. They resemble each other the way a flat-packed cardboard box resembles assembled furniture: same thing in spirit, completely different states. const user = { name : "

How I Cut Costs 65% Migrating LangChain to DeepSeek I want to tell you about a switch I made recently that genuinely surprised me. If you're running LangChain in production and haven't explored the DeepSeek models yet, this one's for you. Let me show you what I learned, what broke, and what I'll never go back to. The short version? I was burning cash on a generic LLM setup. I migrated to DeepSeek through Global API's unified interface, and my monthly inference bill dropped by over 60%. Setup took me less time than brewing coffee. Let me walk you through it. Why I Even Looked at This in the First Place Here's the thing about working in AI engineering: the model landscape moves so fast that whatever you chose six months ago is probably overpriced now. That's been my experience, anyway. When I first built my LangChain pipeline, I defaulted to a popular name-brand model because, well, that's what everyone was using. It worked. It was fine. Then I looked at my AWS bill. That's when I started digging into alternatives. And let me tell you, the rabbit hole is deep. Global API alone exposes 184 AI models at prices ranging from $0.01 to $3.50 per million tokens. That's a wild spread. The trick is finding the sweet spot where cost meets quality, and for migration workloads (think: code translation, schema conversion, content rewrites), I found it with DeepSeek. Let me show you the numbers that actually mattered to me. The Pricing Reality Nobody Talks About I built a comparison table when I was making this decision, and I want to share it because staring at these numbers side by side is what convinced me. Here's the lineup I evaluated through Global API: DeepSeek V4 Flash sits at $0.27 per million input tokens and $1.10 per million output tokens, with a 128K context window. That's my default for most production traffic now. Fast, cheap, and smart enough for almost everything. DeepSeek V4 Pro comes in at $0.55 input and $2.20 output with a beefier 200K context. I use this when

Every developer has done it. You hit a README, you see the install command: curl -fsSL https://example.com/install.sh | sh And you run it. Maybe you skim the script first. Maybe you don't. But you run it. I've been doing this for years. And each time, a small voice in the back of my head says: you have no idea what that script actually does. You just piped a stranger's code straight into your shell. Eventually I got tired of ignoring that voice. What the pattern actually is curl | sh is not a bad pattern — it's a fast, convenient pattern with a real trust gap. The script runs with your permissions, in your shell, right now. It can: Install something with sudo Delete files with rm -rf Write to your disk with dd Access your SSH keys or .env files Set up a cron job or a systemd service that runs again next reboot Decode and run a payload with base64 | eval Most install scripts do none of these things maliciously. But many do several of them legitimately — and you wouldn't know which ones until something went wrong. --- ## What I built instead I'm a solo founder based in Ouagadougou, Burkina Faso. I build with heavy AI pairing — I'm not a trained engineer, I work with Claude, review the output, and ship. This tool ( peek ) was AI-paired and reviewed by me before release. peek is a ~130-line POSIX shell script that sits in front of the pattern: # Instead of: curl -fsSL https://example.com/install.sh | sh # Do: peek https://example.com/install.sh Before anything runs, peek: Fetches the script Scans it for risky patterns Prints a risk score and the exact dangerous lines Asks you to confirm — and refuses to auto-run a HIGH-RISK script unless you type RUN You can also pipe into it, or run it in analysis-only mode: curl -fsSL https://example.com/install.sh | peek # analyze from a pipe peek --print ./downloaded.sh # never runs, analysis only What it flags (and what it doesn't) The patterns peek checks: Root escalation — sudo , running as root Destructive file ops — rm -rf , fi

You've built a checkout reactor that reserves inventory, charges a card, generates a shipping label, and sends a confirmation email. It runs through Sidekiq. When something fails, compensation logic rolls it back. It works. Then your team asks: "How many checkouts failed this week? Which step? How long does the charge step take at p99? Can we see a trace through the entire system?" Before v0.5.0, you'd need to add logging calls to every step, build a custom Sidekiq middleware, and figure out how to correlate traces across async job boundaries. Now it's one line of config. Enter Middlewares Ruby Reactor 0.5.0 introduces a middleware pipeline — the same pattern that powers Rack, but designed for saga execution. A middleware is a plain Ruby object that hooks into the reactor lifecycle: class TimingMiddleware < RubyReactor :: Middleware def initialize ( ** options ) super @started = {} end def on_start_step ( step_name , _arguments , _context ) @started [ step_name ] = Process . clock_gettime ( Process :: CLOCK_MONOTONIC ) end def on_complete_step ( step_name , _result , _context ) started = @started . delete ( step_name ) return unless started elapsed = Process . clock_gettime ( Process :: CLOCK_MONOTONIC ) - started logger . info ( "step #{ step_name } took #{ elapsed . round ( 4 ) } s" ) end end This middleware times every step. Register it globally: RubyReactor . configure do | config | config . middlewares = [ TimingMiddleware ] end Now every reactor — every checkout, every refund, every data import — gets step-level timing, for free. The full lifecycle (20+ events) Middlewares can observe the complete execution lifecycle: Phase Events Reactor on_start_reactor , on_complete_reactor , on_failed_reactor Step on_start_step , on_complete_step , on_failed_step , on_retry_attempt Compensation on_start_compensation , on_complete_compensation , on_failed_compensation Undo on_start_undo , on_complete_undo , on_failed_undo Coordination on_lock_acquired , on_lock_failed , on_

Hey pessoal! 👋 Queria compartilhar um projeto que venho desenvolvendo para resolver um problema comum: encontrar uma biblioteca de ícones limpa, consistente e fácil de integrar, sem peso desnecessário no projeto. Apresento o coreIcons — uma coleção organizada de 352 ícones de desenvolvimento feita para workflows modernos. 📦 Por que o coreIcons? Leve e Rápido: Impacto mínimo no carregamento das suas aplicações. Organizado: Desenvolvido com uma grade (grid) e estilo totalmente consistentes. Focado no Dev: Feito para se encaixar perfeitamente nos seus projetos de frontend ou full-stack. 🚀 Como Usar É super simples! Basta acessar a nossa demonstração ao vivo, navegar pela coleção e clicar em qualquer ícone. O sistema vai fornecer instantaneamente a URL correta ou o snippet do ícone escolhido para você copiar e usar na hora. 🌟 Apoie o Projeto & Conecte-se! Se você achar essa biblioteca útil, por favor, considere deixar uma estrela ⭐️ no nosso repositório do GitHub ! Seu apoio ajuda o projeto a crescer e a alcançar mais desenvolvedores na comunidade. Também quero deixar um agradecimento enorme a todos que apoiam projetos open-source, contribuem com feedbacks e ajudam a construir um ecossistema melhor para todos nós. Vamos construir juntos! 🔗 Acesse o projeto Repositório no GitHub: https://github.com/mauriciospark/coreIcons Demonstração / Site: https://mauriciospark.github.io/coreIcons/ O que você achou? Deixe suas ideias, feedbacks ou sugestões nos comentários abaixo! 👇 Hey everyone! 👋 I wanted to share a project I've been working on to solve a common problem: finding a clean, consistent, and easy-to-integrate icon library without overhead. Meet coreIcons — an organized collection of 352 development icons built for modern workflows. 📦 Why coreIcons? Lightweight & Fast: Minimal footprint for your applications. Organized: Designed with a consistent grid and style. Developer-Centric: Built to fit smoothly into your frontend or full-stack projects. 🚀 How to Use It's extremely

This is a submission for the June Solstice Game Jam What I Built Beat the Oracle is a daily FIFA World Cup 2026 prediction game where you go head-to-head against an AI — Google's Gemini 1.5 Flash — to call match scores before kickoff. Out-predict the machine and you win the day's Turing Test. Lose, and the Oracle has outsmarted you... until tomorrow. Every day you're served the same 5 matches as everyone else: some already played (scored instantly), some upcoming (lock in your call and come back). The Oracle reads each team's recent World Cup form and makes its own prediction with written reasoning — which you only see after you've locked in yours. No peeking, no cheating. Just you versus the machine. Scoring: Result Points Exact scoreline 3 🎯 "Enigma Cracked!" Correct result (W/D/L) 1 ✅ Miss 0 ❌ Why this fits the June Solstice Game Jam This jam asked for a game inspired by the solstice or any June celebration — and Beat the Oracle is stitched to June on two threads the challenge itself calls out: The World Cup is June's global celebration. The prompt names it directly: "the electric teamwork and high stakes of the World Cup, bringing the entire planet together in the spirit of play." That's the playground this game lives in — and as a bonus, I built it from 🇨🇦 Canada, a 2026 host nation . June is Alan Turing's month. Born June 23rd, Turing is the reason this jam has a "father of computing" prize at all. So I didn't bolt a Turing reference onto a football game — I built a playable Turing Test and gave it a World Cup costume. Turing's 1950 question, "Can machines think?" , becomes a question you answer with your gut every single day: can a machine predict football better than you? And the "daily" loop — new matches each day, your score reset, the machines winning "for today" — leans into the solstice's own theme of cycles and the passage of time. Every day is a fresh test. Every match is a new cipher. 🔗 Live demo: hema-nambi.github.io/BeatTheOracle Code Hema-Nambi /

WP Maintenance Manager ships from a single Python codebase to both Windows and macOS. "Python is cross-platform — write once, run anywhere," the saying goes. The reality is that the distribution pipeline is completely separate per OS , each with its own pitfalls. PyInstaller / Inno Setup / Apple Notarization / eSigner — the release cycle is a combination of OS-specific toolchains. Here's the full picture, plus what to watch out for at each step. (The choice of internal architecture, Flask + browser UI, is covered separately in why we built a desktop app on local Flask + browser UI ; this post is about distributing that architecture across two operating systems.) The per-OS pipeline at a glance Step Mac Windows Build PyInstaller ( --target-arch x86_64 ) PyInstaller Distribution format .app bundle → .dmg folder → .exe installer Installer creation hdiutil / create_dmg.sh Inno Setup ( .iss script) Code signing codesign + Developer ID certificate eSigner CSC (cloud signing) Pre-distribution validation Apple Notarization SmartScreen reputation buildup Final artifact WP_Maintenance_Pro_X.X.X.dmg WP_Maintenance_Pro_Setup_X.X.X.exe Both OSes share PyInstaller, but the path diverges from there. Mac sits inside Apple's review process; Windows runs through Microsoft's reputation system. They're fundamentally different ecosystems. Mac — PyInstaller → sign → Notarization → DMG The Intel / Apple Silicon trap The first trap in Mac PyInstaller builds is architecture . Running pip install + python build_app.py on an Apple Silicon Mac without thinking produces native binaries (like cffi ) for arm64 only — which then don't run on Intel Macs at all. The fix is to run the entire build through arch -x86_64 : arch -x86_64 pip3 install -r requirements.txt arch -x86_64 python3 build_app.py That produces an .app containing only x86_64 binaries, which runs natively on Intel Macs and through Rosetta 2 on Apple Silicon — a unified distribution. Sign inside-out The .app PyInstaller produces conta

This post is a TypeScript implementation of the pattern described in "Beyond the Agentic Loop: The Orchestrator Pattern for Multi-Agent Systems" by Amogh Ubale (Stackademic). The original is Python with generic agents; here we keep the idea intact and re-theme it as a shopping assistant so the three execution modes have something concrete to chew on. All the design credit goes to that article — go read it first. The cast: a handful of shopping agents Before the pattern, the scene. The demo is a small storefront assistant backed by a few single-purpose agents: Catalog — list the categories on offer, or search products by keyword and price. Inventory — check stock and availability for a product. Pricing — look up the current price and any active promotions. Reviews — fetch a product's rating and review highlights. Order — place an order for a product. A customer request might need just one of these, several of them at once, or a few in a strict order — and deciding which of those shapes a request calls for is exactly what the orchestrator is for. The problem: the LLM as a while loop The default way to build a multi-agent system is the agentic loop : you hand the model a bag of tools and let it drive. think → call a tool → observe the result → think again → call another tool → … The LLM is both the brain and the control flow. That's wonderfully flexible, and it's the right tool when the task is open-ended and you genuinely don't know the steps in advance. But in production it has three nasty properties: Unpredictable shape. Every "think" step is another LLM round-trip, so a three-agent task might take 3 calls or 9 — you don't know until it runs, and latency swings with it. (The article clocks a representative three-agent query at ~7 calls through the loop; the wall-clock and spend follow, but the unpredictability is the part that actually bites.) Non-determinism. The same question can take a different path each time, which makes behavior hard to reason about and hard t